No OneTemporary
Actions

Size

164 KB

Subscribers

None

View Options

	diff --git a/an-advanced-introduction-to-gnupg.org b/an-advanced-introduction-to-gnupg.org
	index 189312d..d995fbd 100644
	--- a/an-advanced-introduction-to-gnupg.org
	+++ b/an-advanced-introduction-to-gnupg.org
	@@ -1,2895 +1,3009 @@
	# % -- mode: org; ispell-dictionary: "american"; eval: (require 'org-ref); eval:(setq org-latex-pdf-process '("lualatex -interaction nonstopmode -output-directory %o %f" "bibtex %b" "lualatex -interaction nonstopmode -output-directory %o %f" "lualatex -interaction nonstopmode -output-directory %o %f")) --

	# This file requires the org-ref module to be installed. org-ref is
	# available at: https://github.com/jkitchin/org-ref . As of July
	# 2017, it is not in Debian. Given the large number of dependecies
	# (which also don't appear to be in Debian), the best way to install
	# appears to be via melpa. This can be done by executing the
	# following three lines of code (C-x C-e at the end of each) and then
	# installing org-ref. Note: melpa doesn't use TLS to transfer the
	# packages never mind checking signatures!
	#
	# (add-to-list 'package-archives '("melpa" . "http://melpa.org/packages/") t)
	# (package-initialize)
	# (package-list-packages)
	#
	# Cheat sheet:
	# C-c ] Insert a reference.
	# org-ref-find-bad-citations

	#+STARTUP: indent showeverything
	#+latex_header: \usepackage{url}
	#+LaTeX_HEADER: \usepackage[T1]{fontenc}
	#+LaTeX_HEADER: \usepackage{mathpazo}
	#+LaTeX_HEADER: \linespread{1.05}
	#+LaTeX_HEADER: \usepackage[scaled]{helvet}
	#+LaTeX_HEADER: \usepackage{courier}
	#+LaTeX_CLASS: book

	# Only start making lists after 5 levels of nesting.
	#+OPTIONS: H:5

	# Turn off the automatic placement of the TOC so that can insert the
	# copyright text first.
	#+OPTIONS: toc:nil

	#+Title: An Advanced Introduction to GnuPG
	#+AUTHOR: Neal H. Walfield

	#+BEGIN_LaTeX
	\clearpage
	\ \vfill
	#+END_LaTeX

	#+LATEX:\noindent
	Copyright \copy 2017 g10 Code GmbH.

	#+LATEX:\noindent
	This work is licensed under a [[http://creativecommons.org/licenses/by/4.0/][Creative Commons Attribution 4.0
	International License]].

	#+TOC: headlines 2

	# If we don't have at least one part, i.e., some org mode header with
	# a single , then all headers get promoted. That is * become *.
	# This is a formatting disaster. Hence this hack. Perhaps at some
	# point we'll use parts.
	* Main Matter

	** Introduction

	GnuPG is an implementation of the OpenPGP protocol, which is used for
	encryption and authentication.

	GnuPG is used to encrypt email. But, this functionality is not just
	used by individuals to preserve their privacy: political activists
	rely on it to organize their activities, journalists rely on it to
	protect their sources, and lawyers rely on it to protect
	attorney-client conversations. Jason Reich, the director of security
	for BuzzFeed, describes the importance of GnuPG to journalists this
	way: "GPG is part of a balanced breakfast of any reporter, especially
	one who wants to protect their sources, and be able to be reached for
	leaks and things of that
	nature."\nbsp{}cite:walfield2017jason-reich-interview.
	# XXX: This should be Michał Woźniak, but for some reason, when
	# org-mode exports it the characters are lost. They don't show up in
	# org-entities either. Just kill the accents for now.
	And, Michal Wozniak, the Chief Information Security Officer at the
	Organized Crime and Corruption Reporting Project (OCCRP), said, "I do
	strongly believe that had we not been using GnuPG all of this time,
	many of our sources and many of our journalists, would be in danger or
	in jail"\nbsp{}cite:walfield2017michal-wozniak-interview. Cindy Cohn,
	the Executive Director of the Electronic Frontier Foundation, goes
	further, and says that the privacy and security that GnuPG offers
	makes it "one of the core tools that we need if we're going to have
	functioning self-government in the United States or around the
	world"\nbsp{}cite:walfield2017cindy-cohn-interview.

	But, GnuPG is not only used for encrypting email. GnuPG protects the
	software updates of nearly all free software-based operating systems
	including Debian, Ubuntu, Red Hat, and SUSE. Although less common on
	the desktop, these systems power two-thirds of all web
	sites\nbsp{}cite:wikipedia-linux-market-share, and are the dominate
	platforms used in the cloud computing sector. That means that even if
	you don't directly use GnuPG, if you use the Internet, your personal
	data is, in part, being protected by GnuPG.

	And, GnuPG is used for much more. People use it to protect data
	archives, such as, backups. Software distributors sign their software
	with it so that users can verify the integrity of a copy. Software
	developers use it to sign their commits
	cite:gerwitz2012repository-integrity. Organizations like Debian use
	it to secure internal processes, such as making sure that a package
	upload is authorized, that a vote is legitimate, and that a
	resignation is authentic. GnuPG is used to secure Bitcoin wallets.
	And, GnuPG is used to sign documents.

	*** History

	Werner Koch started GnuPG in 1997\nbsp{}cite:koch1997first-release.
	But GnuPG's roots lies in PGP, an encryption program originally
	written by Phil Zimmermann in 1991\nbsp{}cite:wikipedia-pgp.
	Zimmermann was a long-time political activist, and wrote PGP to allow
	activists to securely store messages on BBSs. Although the source
	code for PGP was available, it wasn't free software. Further, due to
	its use of RSA for public-key cryptography, and IDEA for symmetric
	encryption, PGP was patent encumbered.

	Around 1996, Richard Stallman, the founder of the Free Software
	Foundation, started appealing to people to create a free replacement
	for PGP. Koch was inspired by this speech, and began working on g10,
	as he initially called it, which was a reference to the tenth
	/Grundgesetz/ (the tenth article of German the constitution), which
	enshrines the right to private communication in Germany. Since the
	reference was considered to be too obscure even for most Germans, the
	name /GNU Privacy Guard/ or GnuPG, for short, was adopted soon after
	the initial release.

	As of 2017, Koch has continued to work on GnuPG as the lead developer.
	Since its start, the project has remained relatively small in terms of
	the number of contributors. But, it was only in 2012 that Koch found
	himself working alone on GnuPG. Prior to that, the project received
	enough funding to employ a couple of developers. In 2012, however,
	GnuPG had a funding crisis, and Koch was forced to lay off his last
	employee. The funding situation continued to deteriorate, and in 2014
	Koch had to take side jobs unrelated to GnuPG to supplement his
	income. The situation was unsustainable, and Koch nearly gave up.
	But, friends convinced him to give a donation campaign one last shot.
	The response was amazing. Not only did he receive enough money to
	fund himself, but he pulled in 250,000\nbsp{}euros in small donations,
	and Stripe, Facebook, and the Linux Foundation each committed to
	donating about 50,000\nbsp{}euros per year. Along with some partially
	unexpected contracts from the German BSI (the Federal Office for
	Information Security), Koch was able to hire five additional
	developers.

	Since then, development of GnuPG has accelerated, and new features are
	being added on a regular basis. For instance, Koch developed a new
	key discovery protocol called the Web Key Directory
	(WKD)\nbsp{}cite:koch2017wkd, there is a new trust model based on
	TOFU\nbsp{}cite:walfield2016tofu, there is official support for a set
	of Python bindings, and the GnuPG developers are actively contributing
	to Enigmail.

	*** OpenPGP Criticism

	OpenPGP has been widely criticized. There are three main criticisms:
	GnuPG isn't easy to use, GnuPG doesn't support deniability, like Off
	The Record (OTR), and GnuPG doesn't support forward secrecy.

	Respond to:
	https://medium.com/@mshelton/how-to-lose-friends-and-anger-journalists-with-pgp-b5b6d078a315

	Respond a bit more in depth to the Matthew Green blog post:
	https://blog.cryptographyengineering.com/2014/08/13/whats-matter-with-pgp/

	Summarize Filippo articles:
	https://arstechnica.com/security/2016/12/op-ed-im-giving-up-on-pgp/
	https://arstechnica.com/information-technology/2016/12/signal-does-not-replace-pgp/

	**** Usability

	GnuPG is infamous for being hard to use. There is a fair amount of
	truth to this. Nevertheless, the argument can be made that some of
	the difficulties are required to achieve the support that it wants to
	achieve. For instance, it is unavoidable that people who are worried
	about active attackers need to think about authentication.

	**** Deniability

	Deniability (or deniable authentication) is the property that
	participants in a conversation are able to authenticate each other's
	messages, but they cannot later prove this to a third party. In OTR,
	this works by having the participants use a shared key for
	authenticating messages. Thus, if Alice knows that she didn't send a
	given authenticated message, then it must have come from Bob. This
	style of authentication is fundamentally different from digital
	signatures, which provide strong evidence that a particular person
	created or at least endorsed a signed message.

	Why deniability is perhaps not so useful as one might imagine:

	https://debian-administration.org/users/dkg/weblog/104

	**** Forward Sececy

	How important is forward secrecy?


	*** Modern Chat Protocols

	Over the past few years, the amount of activity in the encryption
	space has increased dramatically. One of the catalysts was almost
	certainly the Snowden leaks in June 2013, which not only motivated
	activists to do some work, but also sensitized the public to the
	work's importance. The area that has probably received the most
	attention has been in the end-to-end instant messaging
	space\nbsp{}cite:ermoshina2016e2e-overview. In particular, Signal,
	whose protocol has been adopted by WhatsApp and Google Allo, has
	received very strong endorsements from many prominent members of the
	InfoSec community. In fact, the creators of the Signal protocol,
	Moxie Marlinspike and Trevor Perrin, received the 2017 Levchin Prize
	at the Real World Crypto Symposium for their work on the protocol.

	The first major difference between OpenPGP and Signal is with respect
	to their scope: signal focuses exclusively on real-time communication.
	This narrow focus has a number of advantages in terms of security. In
	particular, because communication is near real time, clients can
	negotiate parameters, and it is possible to implement forward secrecy.

	The other major difference is that OpenPGP focuses on a decentralized
	model whereas these solutions tend to be walled gardens.

	Signal uses the telephone number as a stable identifier, which is a
	strong identifier.

	Unlike GnuPG, these tools focus on real-time communication.

	*** Privacy

	Address nothing to hide argument (that misses the point---everyone
	needs privacy).

	*** Scope

	As its title suggests, this book is intended to be an advanced
	introduction to GnuPG. It is explicitly /not/ a reference manual.
	That is, the focus is not on providing a highly technical, exhaustive
	guide covering exactly what GnuPG does, but on gradually building up
	reader's understanding. This isn't a value judgment; I believe that
	the two are complementary. And, my hope is that after reading this
	book, you'll have a solid understanding of GnuPG's internals, and can
	quickly use GnuPG's reference manual to fill in any required details.

	** A GnuPG Primer

	Examples of how to use gpg from the command line. Cover all of the
	important stuff and little to none of the esoteric options. E.g.,
	generating an online key, encryption, decryption, signing (inline or
	detached, clearsign), verifying sigs, using the ~--edit-key~
	interface. Adding a new user id. Retiring a user id. Revoking a
	key. Signing someone's key. Setting owner trust. To armor or not to
	armor. Talk about importing and exporting keys (including import and
	export filters). Some useful options.

	Listing keys. Talk about the different search methods, e.g.,
	prefixing ~@~ to only search on the email.

	Note that the right way to interact with GPG is not by screen
	scraping, but by using ~gpg~ 's ~--status-fd~ family of options or
	using the GPGME library (or one of the many bindings), which remove
	the need to parse ~--status-fd~ 's output.

	GPG is not a library. Talk about how this arose historically. The
	tension between providing a user interface and a programming API
	(former wants convenience and implication, the latter not.) If you
	want to program GnuPG then it is recommended that you use GPGME (or a
	binding built on top of GPGME). A lower level interface is
	~--status-fd~. Has been around since GnuPG 1.2. Example of why it is
	important to use this interface.

	Groups/aliases

	** Cryptography

	Most readers of this book probably already understand how public-key
	cryptography works. Perhaps not at the mathematical level, but at
	least at the conceptual. But, most readers of this book also need to
	be able to explain public-key cryptography for lay people.

	1. What is cryptography? Basically scrambling a text (using
	permutation and substitution).

	1. Example: most people have probably secured a zip file with a
	password.

	1. How does that work? A simple approach is to imagine that each
	letter is a number---A is 0, B is 1, C is 2, etc.---and then add
	(without carrying---that is B (2) + Z (25) is 27, 27 is larger than
	25, so do: 27 - 26 = 1 and take 1, i.e., do modular 26 arithmetic)
	the plain text to the password. For instance consider the text
	"Meet me in Mantua" and the password "tank boil throw letter".

	MEETMEINMANTUA
	+ TANKBOILTHROWLETTER
	--------------------
	....

	If you know the password, you can easily reverse the process. But
	if you don't know the password, it is effectively impossible to
	recover the plaintext given the ciphertext.

	Note: if the password is at least as long the text and password is
	never reused, this is referred to as a one-time pad and is the
	strong known cryptography.

	1. This approach doesn't scale. If you want to communicate with
	multiple people, you need a remember a password for each person.

	1. Problem solved using public-key cryptography. Instead of sharing a
	password, each person has a so-called public key and a so-called
	private key. Using a public key cryptography, for Romeo to encrypt
	a message to Juliet, he just needs to know her public key. Juliet
	can decrypt the message using her private key. The nice thing
	about the public key is that it can be shared with anyone.

	1. How does it work? Based on so-called one-way puzzles. Consider
	factoring ~221~. To do this, you could try every number from 2 to
	the square root of 221 and see if it evenly divides 221. For 221,
	this doesn't take that long to do, but for a 1000\nbsp{}digit
	number, it could take forever---even for a computer and although
	there are some improvements over to the simply method, none are
	significantly faster. But, if I told you that the factors are 13
	and 17, they you can /verify/ that very quickly. This is basically
	how public-key cryptography works. There are also different
	one-way puzzles.

	1. How to imagine public key encryption? We can think of the public
	key as the blue prints for a safe (or padlock) that anyone can
	build around a message, but once that message is in the safe, it
	can only be opened using the recipients corresponding private key.

	Explain signing.

	Give other examples of how to explain public key cryptography.

	Talk about threat modeling. What are you trying to protect? From
	whom? What resources does the adversary have?

	** OpenPGP

	GnuPG is an implementation of OpenPGP, an encryption standard
	published by the Internet Engineering Task Force (IETF). The IETF's
	main activity is the development and promotion of standards related to
	the Internet. Since its formation in 1986, the IETF has standardized
	many ubiquitous Internet protocols including the HyperText Transfer
	Protocol (HTTP), and the Transport Layer Security (TLS) protocol.
	Each standard is managed by a working group, and anyone can
	participate by joining the appropriate mailing list. The working
	group responsible for OpenPGP is fittingly called /The OpenPGP Working
	Group/.

	OpenPGP consists of three main parts. First, OpenPGP specifies a
	collection of cryptographic algorithms for encrypting and decrypting
	data, generating and verifying digital signatures, and deriving keys
	from passwords (so-called /key derivication functions/ or KDFs).
	These are built on top of more basic cryptographic building blocks
	like SHA-1 (a hash algorithm), AES (a symmetric cipher), and RSA (an
	asymmetric cipher, which is also known as a public-key algorithm).
	For the most part, the specification does not define these algorithms;
	it simply says which algorithms should be used where and how to use
	them. Second, OpenPGP defines a packet-based message format. This
	format is used not only for exchanging encrypted messages, but also
	for transferring keys and key meta-data. Finally, OpenPGP includes
	functionality to help manage keys. This functionality includes the
	ability to revoke a key, and to sign keys.

	The first version of the OpenPGP protocol was published in 1996 as
	RFC\nbsp{}1991. (Although, at that point it was still known as the
	PGP prototcol.) Since then, the protocol has undergone two major
	revisions. The most recent version was published in 2007 as
	RFC\nbsp{}4880. In 2015, the OpenPGP community again reformed the
	OpenPGP working group to update the
	specification\nbsp{}[[cite:openpgp-working-group-charter]].

	The major goals for the next version are: the deprecation of some old
	cryptographic algorithms like SHA-1, the introduction of some new
	cryptographic algorithms based on elliptic curves, the addition of
	modern message integrity protection in the form of something like
	Authenticated Encryption with Associated Data (AEAD), and an updated
	fingerprint format.

	From an application programmer or user's perspective, the working
	group is not considering any major changes to the existing
	functionality; they are primarily tightening the standard's security
	and cleaning up a few issues. This is true even of OpenPGP's use of
	SHA-1, which, although SHA-1 has many flaws, is still considered safe
	in the way that OpenPGP uses it. That is, the changes are mostly to
	proactively---not reactively---address weaknesses. In the words of
	the cryptographer Peter Gutmann, "OpenPGP is still too good enough,
	there's lots of things there that you can nitpick but nothing really
	fatal, or even close to fatal"\nbsp{}cite:gutmann-too-good-enough.

	*** Data at Rest

	OpenPGP is used to protect both data at rest as well as data in
	motion. Whereas data at rest refers to data that is stored, e.g., on
	a hard drive, data in motion refers to data that is transferred, e.g.,
	via HTTP. Thus, an encryption scheme that only protects data in
	motion, such as TLS, removes the encryption on receipt; the data is
	only protected on the wire. Another way to think about the difference
	between data at rest and data in motion is that encryption that
	protects data at rest protects it in time and space whereas encryption
	that protects data in motion only protects it in space. Yet another
	way to think about the difference is that data at rest is to the ~tar~
	or ~zip~ tools as data in motion is to HTTP or XMPP.

	The decision to protect not only data in motion, but also data at rest
	using the same scheme significantly constrains the solution space. In
	particular, because data at rest may be accessed asynchronously with
	respect to the encryption, there is no possibility to negociate
	parameters on the fly.

	Consider an encrypted backup. When you encrypt the data, you can only
	use the strongest encryption that is available at the time of the
	encryption. When you access the data 10 years later, your
	implementation needs to support that now old encryption algorithm;
	there is no way to go back in time and say to your former self,
	"could you use this implementation instead?"

	# I don't think 10 year old backups are that relevant. Typically only
	# the latest snapshot is interesting, which should be fresh. And
	# backups are not longterm preservation archives for many other
	# reasons. (marcus)

	An additional consequence is that upgrading the cryptography becomes
	very difficult. It is not possible to completely deprecate old
	algorithms, because old messages (like our backup) still need to be
	decrypted. Similarly, since people continue to use old software, we
	often cannot use the latest and greatest encryption scheme, because
	they might not be able to decrypt the data!

	Another result of this decision to protect data at rest is that
	enabling forward secrecy is not possible. Forward secrecy is an
	oft-lauded encryption property, which prevents old encrypted messages
	from being decrypted if the private key material is somehow
	compromised. Forward secrecy works by mutating the key material in
	time. This scheme is fine if you never need to decrypt old messages
	(as is typically the case for data transferred via HTTPS, say), but
	doesn't work at all for data at rest: if you want to decrypt some data
	a week later, nevermind 10 years later, then you won't be able to if
	you've destroyed the private key material needed to decrypt it!

	Perfect secrecy becomes even more complicated when a user has multiple
	devices, and all devices should be able to decrypt all messages.
	OpenPGP doesn't require that those devices somehow synchronize their
	state after the private key is copied. But, some type of
	synchronization is necessary for forward secrecy.

	This raises the question: why have a single algorithm for both data in
	motion and data at rest? The reason is that OpenPGP messages are
	often not stored on a trusted host or even processed on a trusted host
	before being stored. Consider email. Email is normally stored on a
	mail server. Even after the mail is read, it remains on the mail
	server so that it can be read later---potentially years later---on a
	different device. Thus, even assuming that we could harden the
	security of the transport layer, it is not clear that when the data is
	on a mail server, it is any less vulnerable than when it is on the
	wire. In fact, data breaches at huge companies entrusted with highly
	personal information from millions or even billions of users, such as
	Yahoo!\nbsp{} and Adult Friend Finder, are evidence that this is not
	the case.

	# Why can't the emails be reencrypted every couple of years to update
	# the algorithms? (marcus)

	*** Unbuffered Message Processing

	OpenPGP is designed to allow unbuffered message processing. This is
	partially achieved by mandating that message packets be sorted
	topologically. That is, if a packet has a dependency, that dependency
	precedes it in the message.

	This property is important for several reasons. First, it allows an
	OpenPGP implementation to run on memory constrainted systems while
	being confident that the implementation can in practice process
	arbitrarily large messages. Second, it ensures that streaming tools
	can be used, e.g., something like ~... \| gpg -e -r key \| ssh ...~.
	Finally, this property helps avoid some denial of service attacks,
	which might otherwise be possible by crafting a malicious message.

	In practice, there are some limitations to the degree to which
	buffering can be avoided. Consider a pipeline in which a message is
	verified, and the output of the message is somehow processed. Because
	the OpenPGP implementation requires the whole message to verify it, to
	process this message in a streaming fashion, the OpenPGP
	implementation has to output the data before it has been verified.
	Now, if the consumer can't process the output in a way that can be
	reverted in the case of a validation failure, the consumer must first
	buffer the data. But, even if it is possible for the consumer to
	recover from a validation failure, it's probably error prone if only
	because code on an error path is rarely tested. Thus, although the
	OpenPGP implementation could avoid buffering data in this situation,
	it has merely shifted the burden.

	Now, there are some more advanced cryptographic constructs, such as
	hash chaining, that make it possible to verify the data bit-by-bit.
	These techniques would help ensure that the consumer only processes
	verified data, which is an improvement over the status quo. But, they
	don't completely solve the problem, because they can't protect against
	message truncation.

	*** OpenPGP Messages

	An OpenPGP message is basically a sequences of packets. OpenPGP
	defines 17 different packet types that are used to not only encrypt
	and sign messages, but also to transfer keys and key signatures or
	certifications, which are used in the web of trust. The format is
	extensible, and this has already been used to add new features.

	An example of a packet type is the symmetrically encrypted data (SED)
	packet. A SED packet contains data that has been encrypted using a
	symmetric algorithm, such as AES. The contents of the packet are zero
	or more OpenPGP packets. That is, OpenPGP messages are nested; a SED
	packet is a container. Typically, a SED contains either a signature
	packet or a compressed data packet, which in turns holds a literal
	data packet, but the specification doesn't impose any limitations.

	This flexibility in message composition is referred to as /agility/.
	It has both advantages and disadvantages.

	A useful advantage that this flexibility offers is that the format can
	be used in unforeseen situations. For instance, the web key directory
	(WKD) uses the non-standard sign+encrypt+sign pattern to facilitate
	spam detection prior to decryption.

	Two important disadvantages of this flexibility are that parsing
	OpenPGP messages is more complicated, and assigning meaning to unusual
	structures can be difficult. As an example of the latter, consider a
	message with two literal data packets, the first of which is signed.
	Assuming the signature is valid, should an implementation report that
	the message is valid? Probably not. The second part could have been
	forged. Alternatively a mail program could show both parts and
	indicate that only the first part is authentic. But, this requires
	educating the user to understand these nuances. Unfortunately
	educating users is known to be extremely difficult.

	*** Encryption

	Most lay people and even many technical people assume that encryption
	includes both an integrity check and authentication. In reality,
	encryption by itself provides neither. This assumption perhaps arises
	due to conditioning from web browsers that not only conflate the two
	concepts, but treat a connection secured with a self-signed
	certificate (which provides encryption, but not authentication), worse
	than those that use neither encryption nor authentication.
	Additionally, in recent years, the term end-to-end encryption has
	entered the mainstream. Although authentication is as important as
	encryption in such systems, only encryption is mentioned. Be that as
	it may, in OpenPGP, encryption and signing are separate, independent
	operations.

	**** Hybrid Encryption

	OpenPGP is a hybrid cryptosystem. A hybrid cryptosystem first
	encrypts data using a symmetric encryption algorithm like AES with a
	random so-called /session key/, and then encrypts the session key
	using the recipient's public key. The result is stored in a so-call
	/public-key encrypted session key/ (PK-ESK) packet.

	There are two important reasons for doing this as well as several
	additional advantages.

	First, public key encryption is thousands of times slower than
	symmetric encryption. Since a session key is just a single block of
	data (which is N\nbsp{}bits for an N\nbsp{}bit RSA key), but the data
	to encrypt could be megabytes or even gigabytes large, this saves a
	lot of processing power.

	Second, it is not unusual to encrypt a message to multiple recipients.
	The most obvious example of this is in the context of email where an
	encrypted email is sent to multiple people. But even in other
	contexts, having multiple recipients is not unusual. Specifically,
	when encrypting data to another party, most programs will also encrypt
	the data to the person doing the encryption so that the data remains
	readable and auditable.

	An advantage of this approach is that it is possible to do
	message-based key escrow. Thus, a company wouldn't need to have
	access to each employee's private key, but whenever the employee
	decrypted an email, the session key could automatically be reencrypted
	with a special escrow key.

	Similarly, if law enforcement forces you to reveal the encryption key
	for some messages, it is sufficient to provide the session keys for
	decrypting the subpoenaed messages. If you had instead provided your
	private key, law enforcement could read any message that had been
	encrypted to you. (In GnuPG, you can extract the session key
	using the ~--show-session-key~ option.)

	Finally, using hybrid encryption, it is possible to encrypt to both
	public keys and passwords. To encrypt a message using a password,
	OpenPGP specifies a key derivation function (S2K), which is used to
	generate a symmetric key. (This is saved in a so-called
	/symmetric-key encrypted session key/ (SK-ESK) packet.) OpenPGP
	allows the symmetric key to be used directly as the session key, but
	it can just as well be used to encrypt a session key. In practice,
	this is primarily interesting to ensure that the sender is able to
	later decrypt the contents of the message by also encrypting the
	session key to her public key.

	**** Algorithm

	Encryption in OpenPGP is a more or less standard hybrid encryption
	scheme:

	1. A random /session key/ is generated.

	1. For each recipient, the OpenPGP implementation encrypts the session
	key using the recipient's public key, and emits a /public-key
	encrypted session key/ (PK-ESK) packet.

	1. If the data should be encrypted using a password, the same thing is
	done, but instead of emitted a PK-ESK packet, a /session-key
	encrypted session key/ (SK-ESK) packet is emitted.

	1. Encrypt the actual data using the session key.

	OpenPGP supports multiple symmetric encryption algorithms. To
	determine which one to use, the OpenPGP implementation selects one
	from the intersection of the recipients' preferred algorithms. This
	information isn't negotiated in real time with the recipients (even
	when this might in theory be possible), but is stored alongside the
	recipient's public key (specifically, in a user ID's self-signature).
	Typically, this is just a list of the algorithms that the OpenPGP
	implementation that generated the key supports at the time the key was
	created, but it can be updated to reflect changes in the
	implementation, and may be customized by expert users. Since all
	implementations are required to at least support TripleDES, and it
	appears implicitly at the end of the list, the intersection is never
	empty.

	**** An Encrypted Message

	To better understand how messages are laid out, the following example
	shows the innards of an encrypted message. This output was created
	using GnuPG's ~--list-packets~ option. ~hot dump~, which is part of
	hOpenPGP, and ~pgpdump~ can do something similar.

	# GNUPGHOME=`pwd`/gnupg/romeo
	#+BEGIN_EXAMPLE
	$ echo 'Let us sojourn in Mantua!' \| \
	> gpg --encrypt -r juliet.capulet@gnupg.net \| \
	> gpg --list-packets
	gpg: encrypted with 2048-bit RSA key, ID C1A010A1D38C4BB8, created 2017-07-07
	"Juliet Capulet <juliet.capulet@gnupg.net>"
	gpg: encrypted with 2048-bit RSA key, ID 5B905AF0423ABB52, created 2017-07-07
	"Romeo Montague <romeo.montague@gnupg.net>"
	# off=0 ctb=85 tag=1 hlen=3 plen=268
	:pubkey enc packet: version 3, algo 1, keyid C1A010A1D38C4BB8
	data: [2046 bits]
	# off=271 ctb=85 tag=1 hlen=3 plen=268
	:pubkey enc packet: version 3, algo 1, keyid 5B905AF0423ABB52
	data: [2046 bits]
	# off=542 ctb=d2 tag=18 hlen=2 plen=85 new-ctb
	:encrypted data packet:
	length: 85
	mdc_method: 2
	# off=563 ctb=a3 tag=8 hlen=1 plen=0 indeterminate
	:compressed packet: algo=2
	# off=565 ctb=cb tag=11 hlen=2 plen=32 new-ctb
	:literal data packet:
	mode b (62), created 1499445579, name="",
	raw data: 26 bytes
	#+END_EXAMPLE

	The example shows a message that Romeo encrypted to Juliet. (Due to
	limitations of the OpenPGP format---OpenPGP only supports timestamps
	between 1970 and 2106---Romeo forward dated the creation time of his
	key.) The first thing that we notice is that even though Romeo only
	specified a single recipient (using the ~-r~ option), the message is
	encrypted to two keys: his and Juliet's. This is because Romeo has
	the ~encrypt-to~ option set in his ~gpg.conf~ file so that he can
	always read messages that he encrypts to someone else.

	***** Packet Metadata

	After listing the recipients, ~gpg~ outputs each packet. Each packet
	starts with a line preceded by a ~#~. This line shows some meta-data
	and the packet's header. Specifically, ~off~ indicates the offset of
	the packet within the stream (this may not be accurate if there are
	compressed packets); ~ctb~ (Content Tag Byte) includes the type of the
	packet, and some information about the length of the packet (if this
	is a new format packet, then ~new-ctb~ will appear towards the end of
	the line); ~tag~ is the type of the packet as extracted from the
	~ctb~; and, ~hlen~ and ~plen~ are the header and body lengths,
	respectively.

	Sometimes the length of a packet is not known apriori. In this case,
	~plen~ will be 0 and ~indeterminate~ or ~partial~ will appear towards
	the end of the line. This can occur when the data is streamed.
	~indeterminate~ means that all data until the end of the message
	belongs to this packet; ~partial~ means the packet uses a chunked
	encoding method to encode the data. The mechanism is similar to
	HTTP's chunked transfer encoding method. These encoding schemes are
	essential for supporting unbuffered operations. See
	Section\nbsp{}4.2.2.4 of RFC\nbsp{}4880 for more details.

	***** The PK-ESK Packets

	The first two packets in the message are PK-ESK packets. Each of
	these holds the session key encrypted to a recipient. A PK-ESK packet
	also includes the 64-bit key ID of key that the session key was
	encrypt to.

	If the key ID wasn't included, then a recipient wouldn't know whether
	a given PK-ESK packet is encrypted with her or someone else's key and
	she would just have to try to decrypt them one by one. The obvious
	consequence is that CPU cycles could be wasted. But, the more
	important reason for avoiding a decryption attempt is that the user
	might have to unlock multiple private keys. This can seriously impact
	an application's usability.

	Avoiding this UX annoyance by including the key ID in the PK-ESK has a
	cost: it leaks meta-data. In practice, however, this information is
	exposed in other places, e.g., at the SMTP level. Nevertheless,
	OpenPGP provides a mechanism to hide this meta-data by setting the key
	ID to ~0~, which means the key ID is speculative. Such key IDs are
	also referred to as wild card key IDs.

	A speculative key ID can be set in GnuPG by either specifying
	~--throw-keyids~ to clear the key ID field for all recipients, or
	~--hidden-recipient~ in place of ~--recipient~ to clear the key ID
	field for a particular recipient.

	***** The Encrypted Data Packet

	Immediately following the PK-ESK packets is an encrypted data packet.
	This ordering is mandatory: it ensures that buffering is not required,
	because the key needed to decrypt the packet is stored prior to the
	data that it decrypts. As already mentioned, an encrypted data packet
	is a container, which contains 0 or more OpenPGP packets. This is not
	obvious from the output of the ~--list-packets~ command, because it
	doesn't show the message's tree structure. In this case, as is
	usually the case, the encrypted data packet contains a single packet.

	In OpenPGP, there are actually two types of encrypted data packets:
	Symmetrically Encrypted Data (SED) packets and Symmetrically Encrypted
	Integrity Protected data (SEIP) packets. Although the former are
	technically allowed by the standard, they are deprecated in practice
	due to security concerns. For instance, it is possible to conduct an
	oracle attack\nbsp{}[[cite:mister2005cfb-attack]], and message extension and
	deletion attacks are also possible. Consequently, when GnuPG
	encounters such a packet, it emits a warning. GnuPG itself will not
	emit an encrypted packet without integrity protection.

	We can see that the encrypted data packet includes integrity
	protection based on the packet's tag (18 instead of 9), and the
	presence of the ~mdc_method~ field in the above output.

	****** Modification Detection Codes

	MDC stands for Modification Detection Code. Like a message
	authentication code (MAC), an MDC can verify a message's integrity.
	But, unlike a MAC, an MDC doesn't say anything about its authenticity.
	A common criticism leveled at the MDC system is that using an HMAC
	would have been better since it is better understood. Ignoring that
	the MDC system has proven to be sufficient for its intended purpose,
	using an HMAC wasn't really an option when the problem was discussed:
	HMACs and MDCs were developed concurrently. (For more historical
	notes, see\nbsp{}[[cite:callas2009mdc]].)

	Prior to the introduction of the MDC system in RFC\nbsp{}4880, it was
	only possible to reliably detect integrity violations using
	signatures. Signatures, however, have the disadvantage that they
	expose the signer's identity, which is sometimes undesirable.

	MDC works by computing the SHA-1 over the clear text and the head of
	the MDC packet. (The rest of the MDC packet is the computed hash.)
	That is, the hash effectively violates the packet framing. But, this
	is exactly the behavior that is required to fully ensure the data's
	integrity: by also including the head of the MDC packet in the hash,
	extension and removal attacks are mitigated. The following example
	illustrates how it works:

	#+BEGIN_EXAMPLE
	+------+-----------------------------------+-------------+
	\| SEIP \| Data (e.g. a literal data packet) \| MDC hash \|
	+------+-----------------------------------+-------------+
	\ / ^
	`--------------------------------------' \|
	SHA-1 -----------------------------'
	#+END_EXAMPLE

	The ~mdc_method~ parameter above seems to suggest that there are
	multiple MDC methods. This is not the case, and was explicitly
	avoided to prevent downgrade and cross-grade attacks; the value of 2
	is simply SHA-1's OpenPGP algorithm identifier. But even though SHA-1
	has since been broken, the relevant security properties for the MDC
	system remain intact. Nevertheless, the working group is considering
	replacing the MDC system with one based on Authenticated Encryption
	with Associated Data (AEAD), which has other useful properties.

	As a final note, the MDC packet is not shown in the output of
	~--list-packets~. This is a technical limitation of GnuPG, which has
	to do with the way the MDC packet is processed. But, given that
	~--list-packets~ is only a debugging interface and not intended for
	programmatic use, this limitation is unlikely to be fixed.

	***** Compressed Packet

	The compressed packet is nested within the encrypted packet.
	RFC\nbsp{}4880 specifies three different compression algorithms---ZIP,
	ZLIB, and BZip2---but notes that they are optional. But even though
	compression is not required, the RFC recommends it as an operationally
	useful (even if not rigorous) form of integrity protection.
	Unfortunately, it has been shown that compressing data prior to
	encryption can enable a chosen plaintext attack as demonstrated by the
	CRIME on TLS, and BREACH on HTTP attacks.

	# https://security.stackexchange.com/questions/39925/breach-a-new-attack-against-http-what-can-be-done/39953#39953

	***** Literal Data

	Nested within the compression packet is a literal data packet. A
	literal data packet contains not only the cleartext, but also a bit of
	metadata. In particular, a literal packet includes a formatting
	field, which indicates whether the contents are binary data or text,
	and, in the latter case, whether the text is believed to be UTF-8
	formatted. The packet also contains a filename, which is helpful when
	transferring a file, but is mostly ignored by GnuPG in practice. And,
	it contains a timestamp. GnuPG sets the timestamp to the current time
	when the packet is created (not the file's ~mtime~).

	It is worth pointing out that when GnuPG is told to decrypt data (~gpg
	--decrypt~), it doesn't look for an encrypted message to decrypt, but
	processes the message and tries to decrypt any encypted data that it
	encounters. This subtle difference in behavior can be important,
	because if GnuPG is told to decrypt a message with just a literal
	packet, it will simply output the contents of the literal packet
	without warning the user that the data was not actually encrypted. If
	a program uses the ability to decrypt a message as an authentication
	check (e.g., in AutoCrypt's Setup Message), this behavior could lead
	to subtle attacks\nbsp{}[[cite:autocrypt-bad-import]].

	*** Signing

	A signature provides cryptographic proof of both the signed data's
	integrity and its authenticity---assuming the key used to sign the
	data is trusted. That is, like a checksum, a signature can be used to
	make sure that the data was not modified in transit. But unlike a
	checksum, a signature can also provide proof of the data's origin (or
	at least, who signed off on the message).

	Note: the exact semantics of a signature are not defined by the
	standard. This is done on purpose, and is viewed by the RFC editors
	as a feature, because, in the end, a signature's meaning is determined
	by the actual human users of the system---some will be more casual,
	and some will be more rigorous no matter what some standard says.

	**** Multiple Signers

	In OpenPGP, it is possible for a single message to include multiple
	signatures created by different keys. This mechanism is useful when
	disparate parties want to sign a document. For instance, multiple
	developers might sign released software. Rather than providing each
	signature separately, it is more useful to combine them into a single
	file.

	In GnuPG, this can be done by specifying each of the keys on the
	command line. For instance:

	#+BEGIN_EXAMPLE
	$ echo 'Good-bye cruel world!' \| gpg -s -u romeo -u juliet
	#+END_EXAMPLE

	A crippling disadvantage of this approach is that all keys must be
	available at the time that the signature is generated, which is rarely
	practical.

	Although OpenPGP's packetized message format makes combining
	signatures relatively easy, GnuPG does not provide support for this.
	Nevertheless, in practice, writing an ad-hoc script is straightforward
	(some hints are
	here:\nbsp{}[[cite:koch2013clearsign-text-document-with-multiple-keys]]).
	And, in the special case that the signatures in question are
	/detached/ signatures, combining them is actually trivial: they just
	need to be concatenated together as shown below:

	#+BEGIN_EXAMPLE
	$ echo 'Romeo and Juliet forever!' > note.txt
	$ gpg --detach-sign -u romeo --output - note.txt > note.txt.romeo.sig
	$ gpg --detach-sign -u juliet --output - note.txt > note.txt.juliet.sig
	$ cat note.txt.romeo.sig note.txt.juliet.sig > note.txt.sig
	$ gpg --verify note.txt.sig note.txt
	gpg: Signature made Tue 11 Jul 2017 11:52:48 AM CEST
	gpg: using RSA key D6636A9EB82A91E94DDEE5066B284A5BE2297415
	gpg: issuer "romeo.montague@gnupg.net"
	gpg: Good signature from "Romeo Montague <romeo.montague@gnupg.net>" [full]
	gpg: Signature made Tue 11 Jul 2017 11:52:59 AM CEST
	gpg: using RSA key E5156E507DCB8D63AC89E5334954FDC67A46B4C5
	gpg: issuer "juliet.capulet@gnupg.net"
	gpg: Good signature from "Juliet Capulet <juliet.capulet@gnupg.net>" [full]
	#+END_EXAMPLE

	In the above examples, the signatures are not nested. That is, they
	are both only over the data, and one could remove either signature
	from the OpenPGP message without impacting the validity of the other
	signature.

	Sometimes, it can be useful to nest signatures. For instance, a
	notary might want to not only notarize some document, but also the
	client's signature over that document. OpenPGP also provides native
	support for this type of signature. In fact, both types can be
	present in the same message. GnuPG does not currently support nested
	signatures.

	**** Algorithm

	As in the encryption case, signing is a two-step process. First, the
	data to be signed is hashed, and then the resulting hash is signed
	using public-key cryptography. This two-step process is primarily
	motivated by performance considerations.

	The exact algorithm that is used is slightly different depending on
	whether the signature should be inline or detached. We start by
	describing how an inline signature is created.

	1. Emit a so-called /One-Pass Signature/ (OPS) packet. An OPS packet
	contains meta-data (what hash algorithm to use, etc.) as well as
	framing information (specifically, whether the signature is nested
	or not).

	1. Hash and emit the data to sign.

	1. Emit a signature packet, which includes the computed hash and the
	signature.

	As its name and the implementation suggest, the OPS packet makes it
	possible to both create a signature, and verify it without buffering
	any data. Since detached signatures are separate from the main
	OpenPGP message, and OPS packets are effectively redundant, to
	generate a detached signature, we just skip the first step. A
	limitation of detached signatures is that they are over the entire
	OpenPGP message. Thus, nesting them is not possible.

	**** Example

	Using our above example with inline signatures, the resulting message
	has the following packets:

	#+BEGIN_EXAMPLE
	$ echo 'Good-bye cruel world!' \
	> \| gpg -s -u romeo -u juliet \| gpg --list-packets
	# off=0 ctb=a3 tag=8 hlen=1 plen=0 indeterminate
	:compressed packet: algo=1
	# off=2 ctb=90 tag=4 hlen=2 plen=13
	:onepass_sig packet: keyid 4954FDC67A46B4C5
	version 3, sigclass 0x00, digest 8, pubkey 1, last=0
	# off=17 ctb=90 tag=4 hlen=2 plen=13
	:onepass_sig packet: keyid 6B284A5BE2297415
	version 3, sigclass 0x00, digest 8, pubkey 1, last=1
	# off=32 ctb=cb tag=11 hlen=2 plen=28 new-ctb
	:literal data packet:
	mode b (62), created 1499772743, name="",
	raw data: 22 bytes
	# off=62 ctb=89 tag=2 hlen=3 plen=333
	:signature packet: algo 1, keyid 6B284A5BE2297415
	version 4, created 1499772743, md5len 0, sigclass 0x00
	digest algo 8, begin of digest 88 56
	hashed subpkt 33 len 21 (issuer fpr v4 D6636A9EB82A91E94DDEE5066B284A5BE2297415)
	hashed subpkt 2 len 4 (sig created 2017-07-11)
	hashed subpkt 28 len 24 (signer's user ID)
	subpkt 16 len 8 (issuer key ID 6B284A5BE2297415)
	data: [2048 bits]
	# off=398 ctb=89 tag=2 hlen=3 plen=333
	:signature packet: algo 1, keyid 4954FDC67A46B4C5
	version 4, created 1499772743, md5len 0, sigclass 0x00
	digest algo 8, begin of digest c5 e3
	hashed subpkt 33 len 21 (issuer fpr v4 E5156E507DCB8D63AC89E5334954FDC67A46B4C5)
	hashed subpkt 2 len 4 (sig created 2017-07-11)
	hashed subpkt 28 len 24 (signer's user ID)
	subpkt 16 len 8 (issuer key ID 4954FDC67A46B4C5)
	data: [2047 bits]
	#+END_EXAMPLE

	***** Compressed Packet

	Again, we see that the message starts with a compression container.
	Since the length of the data is not known apriori, the length is
	marked as ~indeterminate~, which means that the packet includes all of
	the data until the end of the message.

	***** One-Pass Signature Packets

	The next two packets are OPS packets.

	These packets include the hash algorithm that was used to generate the
	signature. This information needs to be available beforehand so that
	the signature can be verified in a streaming fashion. The hash
	algorithm, which is also known as the message digest algorithm, is
	indicated by the ~digest~ field in the output.

	Another piece of information that is necessary to verify the data in a
	streaming manner is how to interpret the data to sign. This is
	determined by the signature's class (~sigclass~). Normally, OPS
	packets are only used with documents (as opposed to keys or user IDs,
	which are so small that buffering isn't an issue). OpenPGP defines
	two types of documents: binary data and text data whose respective
	classes are ~0~ and ~1~. For binary documents, the data is hashed as
	is; for text documents, the OpenPGP implementation first converts line
	endings to ~<CR><LF>~ before hashing.

	The OPS packets also include the signer's key ID and the public key
	algorithm used to generate the signature. This information is
	strictly speaking redundant as it is also stored in the matching
	signature packet, but it can help the implementation identify several
	common cases in which it can't verify the signature prior to actually
	computing the hash. Specifically, the implementation can't verify a
	signature if the signer's public key is unavailable, or the public key
	algorithm used to compute the signature is not supported (even if the
	hash algorithm is supported). In such cases, the implementation can
	fail early, or just skip the hashing, which saves some CPU cycles.

	Finally, OPS packets include framing information. In GnuPG, this is
	referred to as the /last signature/ flag. In the above output, it is
	referred to ~last~. If ~last~ is 1, then the signature is over all of
	the following data up to the OPS's corresponding signature packet; if
	~last~ is 0, then the signature is not nested and is only over the
	data following the next OPS packet with ~last~ equal to ~1~.

	Given this definition of ~last~, we see that the first signature in
	the above example is not nested (~last~ is ~0~), but the second is.
	Thus, both signatures are over the data; the outer signature is /not/
	over the inner signature, just the data.

	To better understand how signatures nest, consider the following
	example, which shows an OpenPGP message with three signatures. The
	first three packets are OPS packets, the middle packet is a literal
	data packet, and the last three packets are the OPS' corresponding
	signature packets.

	#+BEGIN_EXAMPLE
	________________________________________________
	,-----> / \
	+-----------+-----------+-----------+------+---------+---------+---------+
	\| A, last=1 \| B, last=0 \| C, last=1 \| Data \| C's sig \| B's sig \| A's sig \|
	+-----------+-----------+-----------+------+---------+---------+---------+
	\| `----> \____/
	`------------------^
	#+END_EXAMPLE

	Working our way in, we see that ~last~ is set for A's signature.
	Thus, A's signature is over everything immediately following the OPS
	packet up to the matching signature packet. That is, it is over not
	only the data, but also over B and C's signatures. In contrast, in
	B's OPS packet, ~last~ is clear. Thus, B's signature is over
	everything following the next OPS packet with ~last~ set to ~1~, i.e.,
	everything follow C's OPS packet, up to, but not including, the
	signature packet matching C's OPS packet. That is, like C's
	signature, B's signature is only over the literal data packet, not the
	data packet /and/ C's signature.

	***** Literal Data

	The literal data packet contains the document to be signed. Of
	course, if the signatures are nested, then the signature may include
	other data as well.

	***** Signature Packet

	The last two packets are the signature packets that match the OPS
	packets at the start of the message. Like braces in a programming
	language, the first OPS packet matches the last signature packet, and
	the second OPS packet matches the second to last signature packet.

	Except for the nesting information, the signature packet includes
	everything present in the OPS packet as well as some additional
	meta-data, and the actual signature.

	The additional meta-data usually includes a timestamp (the OpenPGP
	Signature Creation Time subpacket), and the user ID that was used to
	make the signature (the OpenPGP Issuer subpacket). There are several
	other pieces of metadata that can be added, but they are not usually
	set in this context.

	The issuer is usually used by a mail user agent to make sure the
	alleged sender matches the signer. For instance, Romeo might have
	verified his father's key, but his father might try to trick him by
	sending him an email that appears to be from Juliet. Because he knows
	that Romeo always checks a signature's validity, he could just sign
	the message with his own key. If the mail user agent only shows
	whether a signature is valid, then Romeo might be tricked. Making
	sure the from header matches the issuer catches this attack.

	*** Keys

	As mentioned above, OpenPGP messages are not only used to transport
	documents, but are also used to transport keys and key signatures.

	In OpenPGP, a so-call /key/ is a lot more than just a public and
	private key pair. Modern OpenPGP keys normally include at least two
	key pairs as well as a fair amount of meta-data.

	**** Multiple Public and Private Key Pairs

	OpenPGP supports multiple key pairs for several reasons.

	First, although it is possible to use the same key pair for encryption
	and signing, if you do, then the act of decrypting a message is
	equivalent to signing it (and vice versa), which could be abused by an
	adversary. In practice, this particular attack is prevented by the
	use of distinguishing padding schemes. But, using separate keys
	avoids this problem and prevents any issues that may be discovered in
	the future.

	Second, having multiple keys makes it possible to largely separate
	identity from key lifetime. In particular, OpenPGP has the concept of
	primary keys and subkeys. The primary key is used to identify the
	OpenPGP key. That is, a key's fingerprint is derived from this key,
	and is independent of any subkeys. This makes it possible for a user
	to revoke individual subkeys without changing her identity. For
	instance, each year you could generate a new encryption and a new
	signing subkey, and revoke the old ones, and there would be no need to
	create new business cards or even inform your contacts that you have
	new keys, because, assuming their software is configured to regularly
	refresh your key, their OpenPGP implementation will automatically find
	the new subkeys since your primary key did not change. In fact, this
	type of key rotation approximates forward
	secrecy\nbsp{}[[cite:brown2001forward-secrecy-for-openpgp]].

	To support an arbitrary number of keys, primary keys and subkeys are
	marked with so-called /capabilities/. There are (perhaps
	surprisingly) four capabilities:

	1. Encryption

	1. Signing

	1. Certification

	1. Authorization

	An encryption capable key can be used for encryption, and a signing
	capable key can be used for signing documents. But, if a key does not
	have the encryption capability, then it should not be used for
	encryption. The certification capability indicates that a key can be
	used for signing /keys/ (as opposed to documents). Thus, since a
	subkey requires a signature to be valid, only a certification-capable
	key can be used to create a new subkey. Finally, the authorization
	capability is used for access control. This is primarily useful for
	using an OpenPGP key with ~ssh~.

	It is entirely possible for a key to have multiple capabilities. As
	mentioned above, it is not advisable to use a key for both signing and
	encryption, but since mathematically certification is just signing, it
	is reasonable to mark a key as both signing and certification capable.

	Whether this is reasonable depends on how the user wants to manage
	keys. For instance, if a signing-capable key is compromised, it is
	possible to recover without generating an entirely new OpenPGP key.
	But, if a certification-capable key is compromised, then the attacker
	effectively owns the identity, and the only way to recover is to
	completely revoke the OpenPGP key and create a new one. This only
	works if users physically separate the certification key from the
	signing key, e.g., by only storing the certification key on an offline
	computer. Since most users don't do this, GnuPG defaults to making
	the primary key both certification capable and signing capable.

	An OpenPGP key can have multiple valid (i.e., not expired and not
	revoked) subkeys with the same capability. In this case, the RFC does
	not specify which subkey should be used; it is up to the
	implementation.

	If there are multiple encryption-capable keys, GnuPG uses the newest
	valid subkey. But this is not the /de facto/ standard. For instance,
	OpenKeychain encrypts a message to all valid encryption-capable keys.

	The OpenKeychain behavior has the advantage that one can store
	different keys on different devices. Then if a particular device is
	compromised, only the subkeys on that device need to be rotated. But,
	operationally, the advantages for encryption-capable subkeys are not
	that large, since an encryption-capable key protects /past/ traffic.
	That is, if an encryption key is compromised, all messages encrypted
	to it are compromised. Thus, a message is compromised if any
	encryption key is compromised. So, in this case, one might as well
	just use a single encryption key.

	This line of logic does not apply to signing-capable keys. If a
	signing-capable subkey is compromised, the attacker can forge
	messages. But, if the user has one signing-capable key per device and
	revokes just the single signing-capable subkey that was compromised,
	then the attacker will be thwarted and only signatures created using
	that key will fail to verify after it has been revoked.

	**** Self Signatures

	As mentioned previously, an OpenPGP fingerprint is derived only from
	the primary key, not the subkeys. This makes sense, since new subkeys
	can be added at any time. Thus, some mechanism is needed to associate
	subkeys with the corresponding primary key. Further, a mechanism is
	needed to associate meta-data with an OpenPGP key. Both of these
	problems are solved using the same mechanism: self-signatures.

	A self-signature is like a normal signature, but instead of being over
	a document, the signature is over structured text, and it is stored
	alongside the OpenPGP key. A self-signature can only be created (or
	rather, is only honored if it was created) by a certification-capable
	key. Since the signature can't be forged, it effectively creates an
	unforgable binding between the OpenPGP key and the data. Thus, to
	determine if a subkey really belongs to a given OpenPGP key, it is
	sufficient to check whether there is a valid self-signature.

	Because OpenPGP packets can be combined in whatever way a user wants,
	an attacker who controls a user's network connection may not be able
	to modify individual packets without detection, but can drop packets.
	Thus, if an attacker has compromised a user's key, the user notices,
	and revokes her key, she is still not safe if the attacker also
	controls the network path, and filters out the revocation certificate
	thereby preventing other users from learning that the key was
	compromised.

	# Shouldn't this last paragraph be in the revocation certificate
	# section? (marcus)

	**** Example

	The following example shows Romeo's key. This key was created by
	GnuPG using the default parameters. Thus, it has a primary key, which
	is signing- and certification-capable, and a single subkey, which is
	encryption capable.

	#+BEGIN_EXAMPLE
	$ gpg --export romeo \| gpg --list-packets
	# off=0 ctb=99 tag=6 hlen=3 plen=269
	:public key packet:
	version 4, algo 1, created 1499443140, expires 0
	pkey[0]: [2048 bits]
	pkey[1]: [17 bits]
	keyid: 6B284A5BE2297415
	# off=272 ctb=b4 tag=13 hlen=2 plen=41
	:user ID packet: "Romeo Montague <romeo.montague@gnupg.net>"
	# off=315 ctb=89 tag=2 hlen=3 plen=340
	:signature packet: algo 1, keyid 6B284A5BE2297415
	version 4, created 1499443140, md5len 0, sigclass 0x13
	digest algo 8, begin of digest 71 f6
	hashed subpkt 33 len 21 (issuer fpr v4 D6636A9EB82A91E94DDEE5066B284A5BE2297415)
	hashed subpkt 2 len 4 (sig created 2017-07-07)
	hashed subpkt 27 len 1 (key flags: 03)
	hashed subpkt 9 len 4 (key expires after 2y0d0h0m)
	hashed subpkt 11 len 4 (pref-sym-algos: 9 8 7 2)
	hashed subpkt 21 len 5 (pref-hash-algos: 8 9 10 11 2)
	hashed subpkt 22 len 3 (pref-zip-algos: 2 3 1)
	hashed subpkt 30 len 1 (features: 01)
	hashed subpkt 23 len 1 (keyserver preferences: 80)
	subpkt 16 len 8 (issuer key ID 6B284A5BE2297415)
	data: [2048 bits]
	# off=658 ctb=b9 tag=14 hlen=3 plen=269
	:public sub key packet:
	version 4, algo 1, created 1499443140, expires 0
	pkey[0]: [2048 bits]
	pkey[1]: [17 bits]
	keyid: 5B905AF0423ABB52
	# off=930 ctb=89 tag=2 hlen=3 plen=310
	:signature packet: algo 1, keyid 6B284A5BE2297415
	version 4, created 1499443140, md5len 0, sigclass 0x18
	digest algo 8, begin of digest 19 f8
	hashed subpkt 33 len 21 (issuer fpr v4 D6636A9EB82A91E94DDEE5066B284A5BE2297415)
	hashed subpkt 2 len 4 (sig created 2017-07-07)
	hashed subpkt 27 len 1 (key flags: 0C)
	subpkt 16 len 8 (issuer key ID 6B284A5BE2297415)
	data: [2043 bits]
	#+END_EXAMPLE

	***** Public Key Packet

	The public key packet normally comes first. It just contains a
	minimum amount of information: the public key algorithm (~algo~), the
	public key parameters (~pkey~), the creation time (~created~), and the
	expiry time (~expires~). Although the ~--list-packets~ output shows
	the key ID, this is not included in the packet; it is shown as a
	matter of convenience. Including it in the packet would be redundant,
	because it is derived from the creation time and the public key
	parameters.

	In the above listing, there is no self-signature for the public-key
	packet. The parameters are, however, protected by the self-signature
	over each user ID packet, which is over not only the user ID packet,
	but also the primary key. It is possible to make signatures just over
	the primary key. But, this is typically only used in the case of key
	revocation.

	Not using a self-signature for the key means that meta-data like user
	preferences needs to be stored someplace else. By convention, they
	are stored in a user ID's self-signature. Consequently, if you have
	multiple user IDs, you could have multiple sets of conflicting
	preferences. This is actually by design: the relevant preferences are
	determined by how the key is addressed, which allows different sets of
	preferences for different environments. So, if you have two user IDs,
	one for work, and one for home, when someone uses your key to encrypt
	to your work email address, the preferences are taken from the work
	user ID. If the caller just specifies the key ID, then the
	preferences are taken from the so-called /primary user ID/. (The
	primary user ID is the user ID with the primary user ID flag set in
	its self-signature. If there are no user IDs that have this flag set
	or multiple user IDs, then RFC\nbsp{}4880 recommends using the user ID
	with the newest self-signature.) Thus, because it is reasonable to
	have different preferences for different user IDs, if the intended
	user ID is known, it---and not the key ID---should be used to address
	the key.

	By convention, self-signatures immediately follow the packet that they
	certify. As such, any direct key signatures would immediately follow
	the public key prior to any user ID or subkey packets. In practice,
	this is not always the case due to implementation bugs or malicious
	intent. Thus, on import, GnuPG will attempt to fix any out-of-order
	packets. This can involve some overhead, but this additional overhead
	is only incurred if the packets are actually out of order.

	When some meta-data is changed, a new self-signature is created.
	Since data that is publish can't easily be deleted, OpenPGP treats the
	key as an append-only log. The result is that a user ID packet, for
	instance, might have multiple self signatures.

	In general, if there are multiple self-signed packets for a given
	packet, only the newest one is used. One important exception is for
	revocation certificates and any designated revoker settings: it is
	necessary to respect these even if a later self signature would
	somehow override them, because this capability could be used by an
	attacker to invalidate a revocation, which would effectively make
	revocations of compromised keys impossible.

	***** User ID Packet

	User IDs are stored between the public key and any subkeys. In this
	example, the key only contains a single user ID.

	A user ID packet just contains a single value: a free-form string. By
	convention (per the RFC), this string is an RFC\nbsp{}2822-style
	mailbox, i.e., a UTF-8 encoded string of the form ~Name
	<email@example.com> (Comment)~.

	Normally, a user ID doesn't require a comment, and, like Romeo's key,
	most keys don't have one. Nevertheless, even though comments can
	(rarely!) be useful for advanced users, it is recommended that most
	tools not offer users the option to set it, because most people don't
	understand what they are for.

	There are two main uses for comments: to distinguish security levels
	and roles. Thus, if a user wants to have two OpenPGP keys associated
	with a given email address, one for low-security communication, which
	is stored directly on the device thereby allowing immediate
	decryption, and one for high security communication, which is, say,
	stored on an air-gapped computer and therefore may introduce a long
	delay if the user is not near the air-gapped computer, comments along
	the lines of "day-to-day key" and "high security key," respectively,
	might be appropriate. Similarly, if a developer has a key that is
	only used for signing commits and releases, a reasonable comment on
	that key could be "dist sig". Daniel Kahn Gillmor takes an even more
	conservative stance, and argues that even these comments are probably
	unnecessary\nbsp{}[[cite:gillmor-user-id-comments]].

	It is also possible to use an image as a user ID. In such cases, the
	image is stored in a so-called user attribute packet. One problem
	with images is that they can be fairly large. Since images like old
	signatures can't be deleted once they are published, and they are
	downloaded whenever a key is retrieved, it is currently recommended
	that images be limited to just a few kilobytes of data.

	Images can be useful since many people are able to more quickly
	associate a person with that person's likeness than with her name.
	Thus, an image could be shown in a Jabber client or a mail user agent.
	However, this should probably only be done for validated keys to avoid
	suggesting authenticity when there is no evidence thereof. Another
	possible use for images is in a graphical depiction of a path in the
	web of trust.

	***** User ID Self Signature

	By convention, the user ID self-signature immediately follows the user
	ID. In addition to binding the user ID to the primary key, it also
	contains additional metadata. As noted above, there may be multiple
	self-signatures, and normally only the newest is used.

	The signature is self-describing. It includes the key that was used
	to create the signature, the algorithm, etc. The ~sigclass~ subpacket
	is ~0x13~, which means that this signature is over a user ID.

	The signature includes a number of hashed subpackets. Hashed
	subpackets are effectively key-value pairs that are validated by the
	signature. The OpenPGP specification includes 22\nbsp{}different
	subpackets including so-called /notation data/, which can be used to
	store arbitrary data. (Notations are describing towards the end of
	this chapter.)

	In this example, there are 10 subpackets. Some of the subpackets
	provide information about the signature itself. This is the case for
	the ~issuer fpr~, ~sig created~ and ~issuer key ID~ subpackets. Some
	of them provide information about the primary key. This is the case
	for the ~key flags~, and ~key expires after~ subpackets. The ~key
	flags~ subpacket is primarily used for indicating the primary key's
	capabilities. The ~key expires after~ subpacket indicates when the
	key expires. An expiration can be extended by creating a new
	self-signature with a later expiration time. Note: the expiration
	time is relative to the key's---not the self-signature's---creation
	time. And, the remaining subpackets describe user and implementation
	preferences. ~pref-sym-algos~, ~pref-hash-algos~, and
	~pref-zip-algos~ specify what symmetric, hash and compression
	algorithms, respectively, the user's OpenPGP implementation supports,
	and the user wants when using this user ID. ~features~ describes what
	advanced features the OpenPGP implementation supports. Currently,
	there is only one flag defined, which indicates that the OpenPGP
	implementation supports the MDC system. And, ~keyserver preferences~
	is a set of flags indicating how the key server should handle the key.

	With the exception of the ~issuer key ID~, all of the subpackets are
	prefixed with ~hashed~. This indicates that this data is part of the
	signed data. Subpackets that are not hashed are considered advisory,
	because an attacker may modify them without detection in transit.

	***** Public Subkey Packet

	The public subkey packets follow the user ID packets. Other than
	their type, these packets are effectively identical to the public key
	packet.

	***** Public Subkey Self Signature

	Like user ID packets, a public subkey packet requires a self-signature
	to validate the key and bind it to the primary key. Typically, a
	subkey packet contains just a few pieces of meta-data, because
	preferences are stored in user ID self signatures.

	There are two minor differences, which are worth pointing out. First,
	whereas the ~sigclass~ field for user ID is ~0x13~, the ~sigclass~ for
	public subkeys is ~0x18~. Second, if the subkey is signing capable,
	then the self-signature must also have a so-called /back signature/ in
	an embedded signature subpacket created by the signing key over the
	primary key and the subkey. Obviously, this back signature should not
	be created for an encryption key based on the aforementioned attacks.

	*** Key Signing

	OpenPGP allows users to validate each other's keys using signatures.
	Thus, if Romeo is convinced that Juliet controls the key
	~0x4954FDC67A46B4C5~, then he could certify it (i.e., sign it) using
	his OpenPGP key. There are two main reasons why Romeo would want to
	certify someone's key.

	First, a certification mechanism of this sort enables the OpenPGP
	implementation to determine whether a key is valid. This information
	is critical when Romeo wants to verify a signed document. In that
	case, Romeo is not just interested in whether the signature is
	mathematically valid, and the data has not be corrupted in transit, but
	he also wants to know whether the signature was really created by
	Juliet. Unfortunately, there is no way for computers to figure this
	out without some help from users. Likewise, when Romeo sends an email
	to Juliet, he wants to be confident that he is really using Juliet's
	key. It is completely possible that Romeo could have a key that
	allegedly belongs to Juliet without realizing it (anyone can create a
	key with any user ID, and upload it to the key servers).

	The other reason that a signature is useful is that it provides a
	mechanism for Romeo's contacts to indirectly verify Juliet's key.
	That is, when Romeo shares this signature with others (e.g., by
	publishing it on a key server), then people who trust him (and this is
	essential!) to validate other people's keys, i.e., to be a so-called
	/trusted introducer/, could use this signature to find a valid key for
	Juliet. The network induced on the signatures is referred to as the
	web of trust although it would be more accurate to refer to is as the
	web of verifications.

	Unfortunately, publishing signatures has the unfortunate side-effect
	of making the user's social graph public. This can have grave
	implications beyond the privacy concerns. For instance, it could be
	used to link a source to a journalist.

	**** Local Signatures

	If a signature shouldn't be published, it is possible to mark it as
	being unexportable. To do this, one would create a local signature.
	This is done in GnuPG by using ~--lsign-key~ instead of ~--sign-key~
	to sign the key. At a technical level, this causes an ~Exportable
	Certification~ subpacket to be included in the signature with the
	value of ~0~.

	Unfortunately, using local signatures is not without problems: it is
	possible to export local signatures and accidentally upload them to a
	key server, and the key server implementations do not automatically
	strip local signatures on import.

	**** Confidence

	When someone verifies a key, she doesn't always have the same degree
	of confidence that the verification is correct. For instance, when
	Romeo signs Juliet's key, he is almost certainly convinced that Juliet
	really controls the stated key. On the other hand, if Romeo is at the
	pub and meets Iago, and he asks him to sign his key, Romeo is almost
	certainly less confident that Iago controls the stated key. This is
	the case even if Iago shows him his government issued identification
	papers. And, it is also the case if he sends an encrypted email to
	the email address in Iago's user ID, and receives a signed reply with
	a shared secret code.

	OpenPGP provides a mechanism for expressing different degrees of
	confidence in the form of three confidence levels ranging from "the
	person said she controls the key" to "I'm confident she controls the
	stated key" as well as a generic, "no comment," level. Other than
	completely ignoring the weakest certification level, this information
	is not included in web of trust calculations by GnuPG. Thus, for all
	intents and purposes, it is just gratuitous meta-data. As such, it is
	better to always use a generic certification
	level\nbsp{}[[cite:gillmor-cert-level]]. This is what GnuPG does by
	default.

	**** Trusted Introducers

	When signing a key, it is possible to indicate that the key holder
	should be a trusted introducer. For instance, an organization may
	have a single key, say ~pgp@company.com~, that they use to sign all of
	their employees' keys. If employees sign ~pgp@company.com~ using a
	trust signature, then anyone who trusts, say, ~alice@company.com~,
	will, as usual, consider ~pgp@company.com~ to be not only verified,
	but, due to the trust signature, a trusted introducer. Consequently,
	that person will also consider any keys that ~pgp@company.com~ signed
	to be verified, which, in this case, is everyone in the company. The
	following example illustrates this idea:

	#+BEGIN_EXAMPLE
	juliet@ alice@ pgp@ bob
	example -- tsign --> company -- tsign --> company -- sign --> @company
	.org .com .com .com
	#+END_EXAMPLE

	In GnuPG, Juliet doesn't actually have to use a trust signature to
	sign ~alice@company.com~'s key: she can just use a normal signature
	and then set the ~ownertrust~ for ~alice@company.com~ appropriately.

	Trust signatures are very powerful and can also be very dangerous. If
	Romeo considers Juliet to be a trusted introducer, and Juliet has
	~tsign~ ed her father's key, then any key that Juliet's father signs
	will be considered verified. Juliet's father could abuse this fact to
	trick Romeo into trusting a key that he forged for Juliet.

	Trust signatures can be constrained. For instance, in the above
	example, Alice probably wants to limit the scope of her trust
	signature of ~pgp@company.com~'s key to just those user IDs associated
	with ~company.com~. To support this, OpenPGP allows a regular
	expression to be associated with a trust signature.

	A trust signature can also make not just immediate connections
	trusted, but also indirect connections. This is extremely dangerous
	and probably only makes sense in very limited situations. For
	instance, in a very large company, each department might have the
	equivalent of the above ~pgp@company.com~ key, and there is a
	company-wide key that ~tsign~ s each department's key. In this case,
	Alice might sign the company-wide key with a depth of ~2~ instead of
	~1~. (When Alice uses a trust level of ~1~, she means that anyone
	that the company verifies is considered verified. A trust level of
	~0~ is equivalent to a normal signature; it doesn't create any trusted
	introducers.)

	In GnuPG, it is currently not easy to modify a signature. For
	instance if you want to convert a normal signature into a trust
	signature, ~gpg~ will complain that the key is already signed. To
	change a signature type or modify a trust signature, it is first
	necessary to revoke the existing signature using the ~revsig~ command
	in the ~--edit-key~ interface.

	**** Non-Revocable Signatures

	Occasionally, it can be useful to make a long-term commitment to a
	signature. This can be done by setting the non-revocable flag. In
	GnuPG, this is done using the ~nrsign~ command in the ~--edit-key~
	interface.

	**** Example

	The following example shows Juliet's key including Romeo's signature
	of her key.

	#+BEGIN_EXAMPLE
	$ gpg --export juliet \| gpg --list-packets
	# off=0 ctb=99 tag=6 hlen=3 plen=269
	:public key packet:
	version 4, algo 1, created 1499443081, expires 0
	pkey[0]: [2048 bits]
	pkey[1]: [17 bits]
	keyid: 4954FDC67A46B4C5
	# off=272 ctb=b4 tag=13 hlen=2 plen=41
	:user ID packet: "Juliet Capulet <juliet.capulet@gnupg.net>"
	# off=315 ctb=89 tag=2 hlen=3 plen=340
	:signature packet: algo 1, keyid 4954FDC67A46B4C5
	version 4, created 1499443081, md5len 0, sigclass 0x13
	digest algo 8, begin of digest 59 1a
	hashed subpkt 33 len 21 (issuer fpr v4 E5156E507DCB8D63AC89E5334954FDC67A46B4C5)
	hashed subpkt 2 len 4 (sig created 2017-07-07)
	hashed subpkt 27 len 1 (key flags: 03)
	hashed subpkt 9 len 4 (key expires after 2y0d0h0m)
	hashed subpkt 11 len 4 (pref-sym-algos: 9 8 7 2)
	hashed subpkt 21 len 5 (pref-hash-algos: 8 9 10 11 2)
	hashed subpkt 22 len 3 (pref-zip-algos: 2 3 1)
	hashed subpkt 30 len 1 (features: 01)
	hashed subpkt 23 len 1 (keyserver preferences: 80)
	subpkt 16 len 8 (issuer key ID 4954FDC67A46B4C5)
	data: [2047 bits]
	# off=658 ctb=89 tag=2 hlen=3 plen=307
	:signature packet: algo 1, keyid 6B284A5BE2297415
	version 4, created 1499445515, md5len 0, sigclass 0x10
	digest algo 8, begin of digest c6 a3
	hashed subpkt 33 len 21 (issuer fpr v4 D6636A9EB82A91E94DDEE5066B284A5BE2297415)
	hashed subpkt 2 len 4 (sig created 2017-07-07)
	subpkt 16 len 8 (issuer key ID 6B284A5BE2297415)
	data: [2046 bits]
	# off=968 ctb=b9 tag=14 hlen=3 plen=269
	:public sub key packet:
	version 4, algo 1, created 1499443081, expires 0
	pkey[0]: [2048 bits]
	pkey[1]: [17 bits]
	keyid: C1A010A1D38C4BB8
	# off=1240 ctb=89 tag=2 hlen=3 plen=310
	:signature packet: algo 1, keyid 4954FDC67A46B4C5
	version 4, created 1499443081, md5len 0, sigclass 0x18
	digest algo 8, begin of digest ee 3f
	hashed subpkt 33 len 21 (issuer fpr v4 E5156E507DCB8D63AC89E5334954FDC67A46B4C5)
	hashed subpkt 2 len 4 (sig created 2017-07-07)
	hashed subpkt 27 len 1 (key flags: 0C)
	subpkt 16 len 8 (issuer key ID 4954FDC67A46B4C5)
	data: [2047 bits]
	#+END_EXAMPLE

	The listing follows the usual format described above. The first
	packet is the public key packet, which is followed by a user ID packet
	and its self signature. And, at the end comes the subkey key and its
	self signature.

	There is one small difference, however. In this listing, Juliet's
	user ID is followed by not one, but two signatures. And, the second
	one is not a self-signature, but Romeo's certification signature: we
	can see from the ~issuer fpr~ subpacket that Romeo, not Juliet,
	created this signature. There are two important things to observe
	here.

	First, Romeo's signature is associated with Juliet's key, not his key.
	Once it is clear that the signature says something about Juliet's key
	and not Romeo's, this makes sense. Nevertheless, many beginners don't
	understand this and think that they somehow own the signature.
	Unfortunately, this arrangement can lead to denial of service attacks.
	For instance, vandals could create many signatures on a particular key
	so that it becomes so large that it can't be imported.

	Second, certification signatures are associated with user IDs and not
	with keys. This avoids bait-and-switch type attacks. Consider Paris
	who convinces Romeo to sign his key. If Romeo signed the key, and not
	the user ID, then Paris could simply revoke the user ID and replace it
	with another, say, Juliet's. Since Romeo would still consider the key
	to be valid, Paris could possibly trick him into believing a message
	from the key is from Juliet.

	*** Revocations

	If a key has been compromised or simply retired, it is essential to
	revoke it so that other people don't accidentally use it. It is also
	important to revoke a user ID if the identity is no longer valid,
	e.g., when leaving an organization, but keeping the same key.
	Occasionally, it can be useful to revoke a user ID certification. For
	instance, you should revoke a certification if: you find out that you
	signed the wrong key; the person who controlled the key somehow lost
	control of it (e.g., he forgot the password, and doesn't have a
	revocation certificate); or, you find out that you signed an
	impostor's key.

	The following example shows what Juliet's key looks like when she
	revokes her own key (the output has been truncated):

	#+BEGIN_EXAMPLE
	$ gpg --gen-revoke juliet \| gpg --import
	...
	$ gpg --export juliet \| gpg --list-packets
	# off=0 ctb=99 tag=6 hlen=3 plen=269
	:public key packet:
	version 4, algo 1, created 1499443081, expires 0
	pkey[0]: [2048 bits]
	pkey[1]: [17 bits]
	keyid: 4954FDC67A46B4C5
	# off=272 ctb=89 tag=2 hlen=3 plen=310
	:signature packet: algo 1, keyid 4954FDC67A46B4C5
	version 4, created 1500052199, md5len 0, sigclass 0x20
	digest algo 8, begin of digest 04 ca
	hashed subpkt 33 len 21 (issuer fpr v4 E5156E507DCB8D63AC89E5334954FDC67A46B4C5)
	hashed subpkt 2 len 4 (sig created 2017-07-14)
	hashed subpkt 29 len 1 (revocation reason 0x02 ())
	subpkt 16 len 8 (issuer key ID 4954FDC67A46B4C5)
	data: [2048 bits]
	# off=585 ctb=b4 tag=13 hlen=2 plen=41
	:user ID packet: "Juliet Capulet <juliet.capulet@gnupg.net>"
	...
	#+END_EXAMPLE

	The revocation is the second packet. It is a self signature on the
	primary key. We know that the packet is a revocation certificate
	based on the ~sigclass~ (~0x20~) as well as the ~revocation reason~
	subpacket. The ~revocation reason~ allows the user to say why the key
	is revoked. Here, the value is ~0x2~, which means that the key was
	compromised. This subpacket can also include a human-readable string.
	In this case, Juliet did not provide any additional information. But,
	in the case that the key is being rotated, it might be helpful to
	include the new key's fingerprint. Of course, this is of limited use,
	since it is not machine readable.

	*** Notations

	RFC\nbsp{}4880 allows signatures to contain arbitrary data. This
	mechanism can be extremely useful for extending the OpenPGP system.
	But, despite its availability, they aren't generally used. One
	example of how they could be used was considered by the Debian
	project, which thought about using notations to store additional
	information about how a developer's identity was
	checked\nbsp{}[[cite:debian-notations]].

	Notations are key value pairs. The key must be of the form
	~key@example.com~. The domain is included to avoid naming conflicts.
	Although the value can be any arbitrary data, GnuPG currently only
	supports free-form strings.

	One limitations of notations is that as they are stored in signature
	subpackets, they must fit into the 64 kilobytes of space available to
	signature subpackets. (Strictly speaking, the hashed area is limited
	to 64 kilobytes of subpackets and the unhashed area has the same
	limitation, but using the unhashed area is not advisable.)

	*** Summary

	This chapter has presented the important details of the OpenPGP
	standard. This introduction wasn't intended for someone who is
	planning to write an OpenPGP parser, but to provide a rough overview
	of the system. Many details have been omitted, as well as several
	minor features (yes, for better or worse, OpenPGP is that feature
	rich). For those looking for more information, the RFC is probably
	the best place to start: it is highly readable, and this introduction
	should hopefully make it easy to navigate.

	** Passwords

	What are passwords used for (symmetric encryption and protecting
	private key). Passwords are not used to protect asymmetric
	encryption. The reason for having a password is to protect the key if
	the device is compromised (e.g., malware or stolen). Thus, a weak
	password does not mean weak transport security; the security of the
	transport is the e.g. RSA encryption. If threat model is typical of a
	private individual, then using a password manager and a relative weak
	password is acceptable.

	How to generate a strong password: need to be able to measure entropy.
	Long passphrase doesn't mean anything: if it is a line from a song, it
	is probably weak. NSA probably tries all of Wikipedia in various
	forms in the first few hours of trying to crack your password. The
	only secure way is to use diceware.

	Snowden: "Assume your adversary is capable of one trillion guesses per
	second." To withstand one year, need 65 bits of entropy! How to
	measure a password's entropy? Need a random password. But that's
	impossible to memorize. Unless we encode it smartly!

	*** Diceware

	Encode using a simple word list

	- /dev/random? 1k words (10-bits entropy per word)
	- dice? $6^4 = 1296$ words (10.3-bits entropy)

	Secure even if adversary knows the word list!

	Examples:
	1. able
	2. about
	3. above

	Required length:
	80 bits = good = 8 words
	120 bits = strong = 12 words

	Examples:
	- percent burst able smash opposite ready blind stab
	- pipe after harm person split seize radar about

	Word lists: Diceware (8k). PGP Biometric word list (512). Voice of
	America's simple English word list (1.5k)

	** Key Generation

	On the surface, generating a key is easy: you just need to call ~gpg
	--gen-key~. For many users, this is appropriate. But, if you have an
	elevated threat model, this is probably not the right approach.

	*** Private Key Management

	- Low security - Online key. Default for --key-gen. (When is this
	okay. Laptop encrypted. If keys were stolen by an unknown,
	unlikely to be cracked. Must trust all local software.)

	- High security - Use a smartcard

	Key stored on a smartcard (GnuK, Nitro, etc.)

	Should use subkeys (Setup slightly more complicated).

	Should store backups on a USB stick, because you can't export private
	key from smartcard. Also what happens if it the SC gets lost.

	Much higher security: Crypto can only be done when key is inserted,
	but, often not obvious what the operation is.

	Note: easier to explain crypto when using a smartcard


	Use tails: Hardened. Wipes memory on shutdown.

	Managing the key:

	- Boot from a USB stick? Medium Security: BIOS might be infected,
	etc.

	- Dedicated offline computer? High security, but still susceptible
	to Bad USB! Old IBM x40 or x60 costs <50 Euros on ebay Remove
	wireless network card!

	*** Key Rotation

	Do set an expiry. (GnuPG defaults to 2 years when creating new keys
	these days.) This guarantees an eventual revocation. People forget
	their passphrase and they don't backup their revocation certificate.
	This is a nice emergency brake.

	Alternative: set a designated revoker. Describe what that means and
	the risks. Show example of how to do it.

	Note: Can easily extend expiration. Expiration bonus: forces people
	to refresh keys.

	Note: When generating a new key, cross sign the keys; revocation
	message is just for humans.

	How to approximate forward secrecy.


	** Validating Keys

	Why is validation important? (What is a MitM attack? Why can't keys
	be validated by a machine?)

	How validation works on the web: x509---centralized and completely
	broken.

	Traditional way to do this in the OpenPGP world is to use WoT.
	Describe how the WoT works. Talk about why it is hard to use: Key
	signing parties are for geeks. Even exchanging fingerprints in person
	is inconvenient.

	Alternatives? If you don't already have the key on a business card,
	just pick up the phone (note: ~--with-icao-spelling~). Talk about why
	using the same medium for getting fingerprint is not good. If you
	want to send an email, then it might be reasonable to use, say,
	twitter direct messages to boot strap a conversation. Both are much
	more secure than no check. How to sign (~--sign-key~ and
	~--lsign-key~).

	Talk about TOFU as an alternative. It's limitation. Nevertheless, in
	practice, probably good. New trust model (since v2.1.10, Dec. 2015).
	Checks identity / key consistency. Model used by ~ssh~. No user
	support required. To enable, add the following to ~gpg.conf~:
	~trust-model tofu+pgp~. Can also set ~tofu-default-policy good~.

	Talk about direct trust and trust always and what they are good for.

	Talk about how to verify: a specific short key ID can be faked in just
	a few seconds. Even a long key ids are not immune to collisions.
	Talk about evil32 / scallion tool.

	*** Key Discovery

	How do you find a key? Traditionally, there are two ways: either via
	a business card or web site or by looking on a key server. The former
	is good, but inconvenient the latter is very, very bad. Key servers
	are not trusted. Anyone can forge a user id, etc. Talk about WKD and
	how it works. Given examples of how to deploy WKD.

	Talk about keybase.io, Autocrypt, and pEp.

	** GnuPG's Architecture

	Talk about GnuPG's split architecture. Explain why it is important to
	separate gpg from gpg-agent, scdaemon, pinentry and dirmngr:
	Components in their own address spaces, which reduces impact of bugs
	(think heartbleed).

	This is different from 1.4.

	GPG is for low security---session encryption, encoding, etc. GPG
	Agent for security operations: password manager, private keys, etc.
	Similar similar to PC and smartcard. (In fact, possible to run
	gpg-agent as a different user id or on a different machine.)

	Smartcard Daemon: Interacts with smartcards (directly or via PC/SC).
	Note: typically packaged separately as ~scdaemon~.

	Pinentry: for interacting with the user (not only passwords, but also
	questions). Started by gpg-agent. Talk about trusted windows and why
	it is important. Several different implementations provide tighter
	integration with a desktop environment. But, these come at a cost to
	security (much more complicated).

	Pinentry fallback if there is no GUI. How to set ~GPG_TTY~ and why it
	is necessary. Talk about different pinentry configuration options and
	what they are good for. In particular, talk about how password
	caching works. Talk about ~gpg-preset-passphraase~ and ~--loopback~
	related stuff in this context. Also talk about ~--keep-tty~ /
	~--keep-display~

	Directory manager is really the network component. Interacts with
	keyservers (HKP, ldap, http) (~--search-keys email@example.com~,
	~--recv-key keyid~). Certificate and CRL cache for gpgsm. Talk about
	different options. In particular, how to best configure ~dirmngr~
	with ~use-tor~, etc.

	Give some details about the different sockets.

	*** gpg-connect-agent

	What it does (communicate with the different components). How to use
	it. Fact that it exposes a command line interface. Use help to
	figure out what to do.

	Show how to script with gpg-connect-agent, e.g., shutting down a
	server.

	*** signals

	Talk about how e.g. SIGUSR1 can be used to cause gpg-agent to dump
	debugging information.

	*** Assuan

	Talk about Assuan. IPC protocol Pipe / socket based. Very simple,
	text-based interface. No interface definition language (IDL). Show
	example of a pinentry session calling ~getpin~.

	Can use ~gpg-connect-agent~ to connect to the running GPG Agent.

	Assuan is a separate package from gpg. Anyone can use it.

	*** Debugging

	Due to the distributed nature of the architecture, it can be hard to
	figure out what went wrong (error messages become more generic as they
	are passed further along the stack).

	~watchgnupg~ helps. Tool for gathering log entries.

	In \texttt{gpg-agent.conf}, add:

	- ~log-file socket:///home/USER/.gnupg/S.log~
	- ~debug-level basic~ (or advanced or expert)

	Run: ~watchgnupg -\:\!-force /home/USER/.gnupg/S.log~

	How to setup a test environment. (Talk about ~GNUPGHOME~ /
	~--homedir~ and where the daemons live.)

	*** configuration

	Talk about gpg.conf, gpg-agent.conf, etc.

	Talk about ~gpgconf~.

	** Good Practices and Tips

	*** Refresh keys.

	When you get a signed message, fetch the key.

	Refresh keys regularly. Why? New preferences. Revocation
	certificates. WoT updates.

	Note: Don't use ~gpg --refresh-keys~. Install parcimonie. Uses tor.
	Random intervals between each key refresh reduces chance of targetted
	attacks and leaking who you are sending messages to.

	*** Key Disclosure

	What to do if you have to disclose the encryption key for a message?
	Don't disclose your private key! This allows decryption of all
	messages. Just disclose the session key. Show example of
	~--show-session-key~.

	*** Backups

	Don't backup the RNG's seed! Exclude ~.gnupg/random_seed~ from backups!

	*** ssh

	Keys Instead of Passwords. Using keys means password is not sent to
	server. Ever enter password for a different server? You've just
	disclosed your password!


	OpenSSH stores private keys on hard drive. Keys are protected by a
	passphrase. Passphrase is cached by ssh agent.

	GnuPG implements the ssh agent protocol. GnuPG can use keys stored on
	a smart card.

	GnuPG's ssh agent: configuration:

	Set ~SSH_AUTH_SOCK~ in ~.bashrc~:

	#+BEGIN_EXAMPLE
	export SSH_AUTH_SOCK=$HOME/.gnupg/S.gpg-agent.ssh
	#+END_EXAMPLE

	Add ~enable-ssh-support~ to ~~.gnupg/gpg-agent.conf~. Restart gpg
	agent. Add public key to ~.ssh/authorized_keys~ file. public key
	obtained by doing:

	#+BEGIN_EXAMPLE
	$ ssh-add -L
	ssh-rsa AAAAB3NzaC1...zyt cardno:000603016636
	#+END_EXAMPLE

	*** Remote gpg-agent

	gpg can use a remote gpg-agent. Running on another computer or as a
	different user.

	- Create a new user, ~gpg~
	- On secure pc, add the following to ~.gnupg/gpg-agent.conf~:

	#+BEGIN_EXAMPLE
	extra-socket /home/gpg/.gnupg/S.gpg-agent-remote
	#+END_EXAMPLE

	On insecure pc, run the following to forward the port:

	#+BEGIN_EXAMPLE
	$ ssh -f -o ExitOnForwardFailure=yes -o StreamLocalBindUnlink=yes \
	> -L /home/neal/.gnupg/S.gpg-agent:/home/gpg/.gnupg/S.gpg-agent-remote
	> gpg@localhost bash -c '{ while sleep 5; do echo NOP; done } \| gpg-connect-agent'
	#+END_EXAMPLE

	Requires at least version 6.7 OpenSSH, which supports forwarding Unix
	Domain Sockets.

	Note: If forwarding fails, exit. If the socket to be forwarded
	already exists.

	Forwards the file ~.../S.gpg-agent~ on the insecure host to the file
	~.../S.gpg-agent-remote~ on the secure host.

	Note: ssh won't expand tildes.

	Loop keeps connection opened and port forwarded. (Could also use
	~autossh~.) Exits when gpg-agent exits.

	** MUA Integration

	This chapter contains guidelines on integrating GnuPG into a mail user
	agent (MUA). Other good sources of information on this topic are
	existing MUAs, in particular, KMail and Enigmail, which probably have
	the best GnuPG integration. This is not to say that our
	recommendations or what KMail and Enigmail implement are optimal. Far
	from it. A common criticism of GnuPG is how difficult it is to use.
	We acknowledge these criticisms, and we particularly welcome help in
	this area. Nevertheless, we suspect that some of the user
	interactions cannot be significantly simplified without compromising
	the security of the system, which is designed to protect the user from
	an active adversary.

	Most people do not have active adversaries. This is particularly true
	in democratic countries. People who live in these places primarily
	turn to a technology like GnuPG to protect their privacy, thwart
	phishing excursions, or fight mass surveillance. These users do not
	have the same security requirements as journalists, activists, or
	lawyers operating in regimes where civil rights are not respected, and
	a single unencrypted message can result in jail time, or worse.

	Given these different classes of users, it is entirely reasonable to
	simplify some of the proposed interaction patterns for those who are
	only interested in protecting their privacy by using encryption
	opportunistically\nbsp{}cite:rfc7435. This is precisely what the
	Autocrypt project is trying to accomplish. Their hope is that trading
	protection from active adversaries for increased ease of use will
	result in greater adoption of encrypted email by people looking to
	protect their privacy, and fight mass surveillance, but don't want to
	be bothered with security issues\nbsp{}cite:autocrypt.

	-These changes need to be made carefully. People currently associate
	-GnuPG and related tools as providing high levels of protection, and
	-may assume that because these new interfaces use GnuPG that they
	-provide the same level of protection. As such, we recommend the MUA
	-make clear to users what level of protection the interface can offer.
	-This could be done using a warning, but text that resembles an EULA is
	-unlikely to be read\nbsp{}cite:boehme10trained-to-accept. Another
	-approach to this problem is to ask the user to select the most
	-appropriate threat model that describes their situation, and then
	-adjust defaults accordingly. This is the approach that the Tor
	-Browser Bundle takes. This has the added benefit of causing users to
	-think about risk assessment.
	-
	-In the GnuPG context, three profiles appear to be reasonable. A very
	-high risk profile for those who fear attacks from a nation state
	-adversary. These users should almost certainly use a security token
	-(which the MUA should help them configure, if necessary), HTML should
	-be disabled, and all dangerous operations should require explicit
	-confirmation. A high risk profile for those who worry about targeted
	-attacks. These users should have to confirm sending unencrypted mails
	-when keys appear to be available, and using unverified keys should
	-require confirmation. And a low risk profile for those who want to
	-protect their privacy. With few exceptions,this profile should avoid
	-interacting with the user. One exception is when the user follows up
	-to an encrypted email, but the reply won't be sent in an encrypted
	-manner.
	-
	-The trade offs are straightforward: someone who requires more security
	-is more sensitive to a mistake, and is more willing to interact with
	-the system to ensure this security. For people who have lower
	-security requirements, not only are these interactions annoying, they
	-can actully hurt security elsewhere: showing dialog boxes that are
	-simply clicked away results in
	+Simplifying user interactions needs to be done carefully. People
	+currently associate GnuPG and related tools as providing high levels
	+of protection, and may assume that because these new interfaces use
	+GnuPG that they provide the same level of protection. As such, we
	+recommend the MUA make clear to users what level of protection the
	+interface can offer. This could be done using a warning, but text
	+that resembles an EULA is unlikely to be
	+read\nbsp{}cite:boehme10trained-to-accept. Another approach to this
	+problem is to ask the user to select the most appropriate threat model
	+that describes their situation, and then adjust defaults accordingly.
	+(If the MUA does not support the user's threat model, then the user
	+should be warned.) This is the approach that the Tor Browser Bundle
	+takes. This has the added benefit of causing users to think about
	+risk assessment.
	+
	+In the GnuPG context, three profiles appear to be reasonable. There
	+should be a very high risk profile for users who fear attacks from a
	+nation state adversary. These users should almost certainly use a
	+security token (which the MUA should help them configure, if
	+necessary), HTML should be disabled, and all potentially dangerous
	+operations should require explicit confirmation. There should be a
	+high risk profile for those who worry about targeted attacks. These
	+users should have to confirm sending unencrypted mails when keys
	+appear to be available, and using unverified keys should require
	+confirmation. And, there should be a low risk profile for those who
	+want to protect their privacy. With few exceptions, this profile
	+should avoid interrupting the user. One exception is when the user
	+follows up to an encrypted email, but the reply won't be sent in an
	+encrypted manner.
	+
	+The trade off these profiles make is straightforward: someone who
	+requires more security is more sensitive to a mistake, and is more
	+willing to interact with the system to ensure this security. For
	+people who have lower security requirements, not only are these
	+interactions annoying, they can actually hurt security elsewhere:
	+showing dialog boxes that are simply clicked away results in
	habituation\nbsp{}cite:boehme11security-cost-of-cheap-user-interaction,boehme10trained-to-accept.

	Communication, of course, necessarily involves multiple parties.
	Thus, if a user with high security requirements communicates with a
	-user with low security requirements, the more casual user could
	-accidentally compromise the more careful user by forgetting to encrypt
	-an email. Thus, consistent with the do-no-harm principle, it is
	+user with low security requirements, the casual user could
	+accidentally compromise the careful user by forgetting to encrypt an
	+email. Thus, consistent with the do-no-harm principle, it is
	important that even an implementation designed for users with low
	security requirements not be too lax.

	*** Integration

	There are two basic ways to add GnuPG support to a MUA: it can be
	added natively, or it can be added via a plug-in. KMail, and Claws
	are examples of MUAs that have native GnuPG support; Enigmail,
	GPGTools, and gpgol are examples of plug-ins.

	One approach isn't necessarily better than the other. But, the
	development of plug-ins tends to be highly divorced from the actual
	development of the MUA with the practical result that the needs of the
	plug-in are often not sufficiently taken into account by the MUA
	developers. This has been a problem for the Enigmail developers, for
	instance.

	One common problem is controlling how messages are rendered: the GnuPG
	support code needs a lot of control over this. This control is
	-necessary to prevent a mimicry attack. Specifically, it is necessary
	+necessary to prevent mimicry attacks. For instance, it is necessary
	to not only show when a message is verified, but also prevent an
	attacker from crafting a message that appears to be verified. One way
	to accomplish this is to style not only the message, but also the
	chrome around the message.

	-The laundry list of things that need to be added to a MUA for
	-reasonable GnuPG support is not long: there needs to be a way to
	-create a key, encrypt messages, verify messages, and do some basic key
	-management. But, all of these things have numerous gotchas that can
	-negatively impact both the user experience, and the security of the
	-system. The point of this chapter is to point out most of these
	-issues to avoid making developers---or worse, their users---rediscover
	-these problems the hard way.
	+The things that need to be added to a MUA for reasonable GnuPG support
	+is not long: there needs to be a way to create a key, encrypt
	+messages, verify messages, and do some basic key management. But, all
	+of these things have numerous gotchas that can negatively impact both
	+the user experience, and the security of the system. The point of
	+this chapter is to point out these issues to avoid making
	+developers---or worse, their users---rediscover these problems the
	+hard way.

	*** Key Creation

	-When a GnuPG-enabled MUA is started, it seems logical to prompt the
	-user to create or import a key if the user has not yet done this.
	-This behavior is reasonable if the user has explicitly enabled GnuPG
	-support by installing a plug-in. However, if the MUA has native GnuPG
	-support, and it is not certain that all users want GnuPG support, it
	-may be best to wait to avoid overwhelming users during the initial
	-setup.
	+When a GnuPG-enabled MUA is started, it would seem logical to prompt
	+the user to create or import a key if the user has not already done
	+so. This behavior is reasonable if the user has explicitly enabled
	+GnuPG support by installing a plug-in. However, if the MUA has native
	+GnuPG support, and it is not certain that all users want to use GnuPG,
	+it may be best to wait to avoid overwhelming the user during the
	+initial setup.

	If a key is not generated immediately, this doesn't mean that the
	-GnuPG support should somehow be hidden or disabled. Even without a
	-key, it is still possible to verify signatures, and show unsigned
	-messages as being insecure. Then, if a user clicks on such a security
	-notice, the MUA can explain why the message is considered insecure,
	-and provide an option for the user to configure GnuPG support.
	-Similarly, it is reasonable to present an option to encrypt a message
	-before a key has been created. If the user selects this option, and
	-there is no key associated with the sending email address, then the
	-MUA can show the key creation wizard. This significantly improves
	-discoverability.
	+GnuPG-related functionality should somehow be hidden or disabled.
	+Even without a key, it is still possible to verify signatures, and
	+show unsigned messages as being insecure. Then, if a user clicks on
	+such a security notice, the MUA can explain why the message is
	+considered insecure, and provide an option for the user to configure
	+the GnuPG support. Similarly, it is reasonable to present an option to
	+encrypt a message before a key has been created. If the user selects
	+this option, and there is no key associated with the sending email
	+address, then the MUA should show the key creation wizard. This
	+significantly improves discoverability.

	The key generation wizard should not only allow the user to generate a
	new key, but also provide an option to import an existing one. When
	the user enters or selects a user\nbsp{}ID, the wizard should look for
	-an existing key with that email address both using WKD, and on the key
	-servers. If there is a matching key, the wizard should ask the user
	-if she wants to import the key or really create a new one. Importing
	-the key might not be possible if the key is a fake, or if the user
	-lost access to the key, e.g., by formatting the computer, or
	-forgetting the key's passphrase. Both are unfortunately extremely
	-common for novice users.
	-
	-When the key generation wizard starts, set the user\nbsp{}ID should
	+an existing key with that email address both in the appropriate WKD,
	+and on any configured key servers. If there is a matching key, the
	+wizard should ask the user if she wants to import the key or really
	+create a new one. Importing the key might not be possible if the key
	+is a fake, or if the user lost access to the key, e.g., by formatting
	+the computer, or forgetting the key's passphrase. Both are
	+unfortunately rather common for novice users.
	+
	+When the key generation wizard starts, the user\nbsp{}ID should
	default to the current identity. For instance, if the user has the
	-email addresses ~alice@posteo.de~ and ~alice@gnupg.net~, and, while
	-composing an email from ~alice@gnupg.net~, clicks on encrypt, the
	-wizard should default to creating a key for ~alice@gnupg.net~. If
	-Alice selects a different identity, then the wizard should explain why
	-the key won't be usable with the email that she is currently working
	-on.
	+email addresses ~alice@posteo.de~ and ~alice@gnupg.net~, and clicks on
	+encrypt while composing an email from ~alice@gnupg.net~, the wizard
	+should default to creating a key for ~alice@gnupg.net~. If Alice
	+selects a different identity, then the wizard should explain why the
	+key won't be usable for the email she is currently composing.

	If the user already has a key, but not one for the current identity,
	it is reasonable for the key creation wizard to offer to add the
	-identity to that key. However, current thinking in the GnuPG project
	-is that users require less training when there is a one-to-one and
	-onto mapping of keys and email addresses than when multiple email
	-address are associated with a single key. For instance, if the MUA
	-offers to add the user\nbsp{}ID to an existing key, it becomes
	+identity to the existing key. However, current thinking in the GnuPG
	+project is that users require less training when there is a one-to-one
	+and onto mapping of keys and email addresses than when multiple
	+user\nbsp{}IDs are associated with a single key. For instance, if the
	+MUA offers to add the user\nbsp{}ID to an existing key, it becomes
	necessary to explain why this might be undesirable, e.g., most people
	probably want separate keys for their private, and their work email.
	-And, later it will become necessary to distinguish between revoking
	-keys and revoking user\nbsp{}IDs, when the user retires her email
	-address. Of course, since many users do use keys with multiple
	-user\nbsp{}IDs, it is necessary for the MUA to support such keys, and
	-explain their meaning when signing keys, for instance.
	+And later, if the user retires her email address, it will become
	+necessary to explain the difference between revoking the key and
	+revoking a user\nbsp{}ID. Of course, since many users do use keys
	+with multiple user\nbsp{}IDs, it is necessary for the MUA to support
	+such keys, and explain their meaning when signing keys, for instance.

	The key generation wizard should make key creation as easy as possible
	by prompting the user for as little information as reasonable. In
	particular, the user should /not/ have the option to enter a comment;
	adding a comment is almost always
	inappropriate\nbsp{}cite:gillmor-user-id-comments. Likewise, key
	-generation parameters including the key length should not be shown.
	-But, the user should be allowed to choose whether the key is published
	-on the Internet. This requires an explanation, which can be made by
	-simile: publishing a key on the Internet is like publishing a
	-telephone number in a phone book, and no one is checking the submitted
	-entries.
	+generation parameters should not be configurable. But, the user
	+should be allowed to choose whether the key is published on the
	+Internet. This requires an explanation, which can be made by simile:
	+publishing a key on the Internet is like publishing a telephone number
	+in a phone book, and no one is checking the submitted entries.

	If it is deemed absolutely necessary that the user be able to tweak
	-key parameters, then they should be hidden unless the user explicitly
	-enables some sort of expert mode: for the most part changing these
	-parameters doesn't actually improve the overall security. For
	-instance, using a 2048-bit RSA key is currently considered
	-sufficiently secure by multiple authorities\nbsp{}cite:keylength. If
	-more security is really needed, then the user should start with the
	-weakest defense, which is almost certainly the user's opsec, not the
	-cryptography. Bruce Schneier, for instance, argues that the Snowden
	-leaks provide strong evidence that the NSA has not broken strong
	-cryptography. Instead, they get the information they want by
	-compromising infrastructure and
	-endpoints\nbsp{}cite:schneier2013staying-secure. The easiest and
	-probably most effective measure is to use a smartcard instead of
	-storing the private key material on the computer.
	+key parameters, then the options should be hidden unless the user
	+explicitly enables some sort of expert mode. The reason is simple:
	+for the most part changing these parameters doesn't actually improve
	+the overall security. For instance, using a 2048-bit RSA key is
	+currently considered sufficiently secure by multiple
	+authorities\nbsp{}cite:keylength. If more security is really needed,
	+then the user should start by improving their weakest defense, which
	+is almost certainly their opsec and not the cryptography. Bruce
	+Schneier, for instance, argues that the Snowden leaks provide strong
	+evidence that the NSA has not broken strong cryptography. Instead,
	+the NSA appears to get the information they want by compromising
	+infrastructure and endpoints\nbsp{}cite:schneier2013staying-secure.
	+The easiest and probably most effective measure is to use a smartcard
	+instead of storing the private key material on the computer.

	There are also practical reasons for not using an overly large key.
	Perhaps the most important one is simply based on performance: it does
	-not take twice as long to verify a signature generated by a 4096-bit
	+not take twice as long to verify a signature generated with a 4096-bit
	RSA key than one generated with a 2048-bit RSA key, but about an order
	of magnitude longer. This performance penalty becomes particularly
	noticeable for 16,384-bit keys.

	**** Revocation Certificate

	-After creating the key, the wizard should prompt the user to save the
	-key's revocation certificate, or offer to print it out (or both!). It
	-is also reasonable to send the revocation certificate to the user in
	-an email. This is the easiest way to make sure the revocation
	-certificate is stored in multiple places, but it has the disadvantage
	-that it gives anyone who can access the user's mail the power to
	-revoke her key. This is problematic, but it is not disastrous: that
	+After creating a key, the wizard should prompt the user to save the
	+key's revocation certificate, or offer to print it out (or both!).
	+For users with low security requirements, it is also reasonable to
	+send the revocation certificate to the user in an email (along with an
	+explanation of what a revocation certificate is, and how to publish
	+it). This is the easiest way to make sure the revocation certificate
	+is stored in multiple places, but it has the disadvantage that it
	+gives anyone who can access the user's mail the power to revoke her
	+key. This weakness is problematic, but it is not disastrous: that
	person would be able to perform a denial of service attack (other
	people would no longer be able to send encrypted messages to the user,
	and signatures generated by the key would no longer be considered
	valid), but could not assume the user's identity, or read encrypted
	-messages. And, creating a new key is straightforward. So, the damage
	-is limited.
	+messages. And, creating a new key is straightforward. So, the
	+potential damage is limited, and for most users probably represents a
	+net win given the benefits of being able to revoke a lost or
	+inaccessible key.

	*** Send Mailing

	The mail composition window should have a toggle to "secure" or
	"encrypt" the current message. When active, this toggle should
	actually cause the message to be encrypted /and/ signed. There should
	-not be a separate toggle for signing the message. As explained
	+/not/ be a separate toggle for signing the message. As explained
	previously, most users assume that encrypting includes signing, and
	don't understand signing at all.

	The button may have a menu that becomes visible after, for instance, a
	long press, which allows the user to select between "Encrypt and
	Sign", "Sign-only", "Encrypt-only" and "No protection." However this
	menu is activated, it should be reserved for advanced users: needing
	-to only encrypt or only sign a message is relatively specialized;
	-normal users should only have to choose between a secure, and an
	-insecure option.
	+to only encrypt or only sign a message is relatively specialized, and
	+these users can be expected to have had training; normal users should
	+only have to choose between a secure, and an insecure option.

	-In some cases, it may make sense to always enable encryption by
	-default, and then require that the user explicitly disable it if
	-encryption is not desired. This avoids mistakenly sending a message
	-unencrypted when it should have been encrypted. However, this default
	-can be annoying for users who do not normally encrypt their mail.
	+For users with high security requirements, it makes sense to always
	+enable encryption by default, and then require that the user
	+explicitly disable it if encryption is not desired. This avoids
	+mistakenly sending a message unencrypted when it should have been
	+encrypted. However, this default can be annoying for users who do not
	+normally encrypt their mail.

	As mentioned earlier, a MUA can deal with this dilemma by setting
	appropriate defaults for the user's threat model. But even for low
	security users, there are cases in which it is clear that encryption
	should be enabled by default. For instance, if the user is replying
	to an encrypted message, then encryption should be enabled. In fact,
	if the user tries to disable encryption, it is reasonable to show a
	warning of the form: "you are replying to an encrypted message, do you
	-really want to display encryption for your reply?" Similarly, if the
	+really want to disable encryption for your reply?" Similarly, if the
	user has recently sent an encrypted mail to a particular user, or
	there is a verified key available for the user, then encryption should
	be turned on. These suggestions are also consistent with a do-no-harm
	mentality.

	-An appropriate default can be more difficult when there are multiple
	-recipients. For instance, when a user replies to an encrypted
	-message, she might not have keys for all of the recipients. In this
	-case, the application should help the user find the keys. In this
	+An appropriate default can be more difficult to find when there are
	+multiple recipients. For instance, when a user replies to an
	+encrypted message, she might not have keys for all of the recipients.
	+But, the application can help the user find the keys, and, in this
	case, finding appropriate keys is actually straightforward: due to the
	-way that OpenPGP encrypts data, the long key\nbsp{}ID of most of the
	-users will normally be embedded in the message (specifically, in the
	-PK-ESK packets). Unfortunately, the key\nbsp{}IDs are subject to
	-tampering, but they are probably more reliable than simply searching a
	-key server for a particular email address. It is also possible to try
	-and find the key using WKS, which provides a basic verification check.
	-In contrast, using a key found on a key server may cause more problems
	-than it solves: the messages may be encrypted, but the intended
	-recipient can't decrypt the message. Making decryption unreliable is
	-a sure way to make people more conservative about when they use it.
	+way that OpenPGP encrypts data, the long key\nbsp{}ID of the sender
	+and any recipients will normally be embedded in the message
	+(specifically, in the PK-ESK packets). Unfortunately, the
	+key\nbsp{}IDs are subject to tampering, but since this requires a more
	+determined adversary, they are more almost certainly much more
	+reliable than simply searching a key server for keys with a particular
	+email address. It is also possible to try and find the key using WKS,
	+which provides a basic verification check. Another reason to avoid
	+key servers is that using a key found on a key server may cause more
	+problems than it solves: the message may be encrypted, but because it
	+is the wrong key, the intended recipient can't decrypt it. Making
	+decryption unreliable is a sure way to discourage the use of
	+encryption.

	Sometimes mails include keys as attachments, or references to them.
	In such cases, the MUA should either import them automatically or
	-provide a button. The keys should always be imported if they are
	-already available locally: the keys might contain updates, such as new
	-subkeys or a revocation certificate. This topic is discussed further
	-in the later section on key discovery.
	+provide a button to allow the user to import them. But, the keys
	+should always be imported if they are already available locally: the
	+keys might contain updates, such as new subkeys or a revocation
	+certificate. This topic is discussed further in the later section on
	+key discovery.

	-**** Keys
	+**** Encryption Keys

	To make it clear whether there is a key for a particular recipient,
	the MUA should add a small icon, e.g., a padlock, next to each email
	address. As usual, to improve discoverability, and provide a reminder
	to encrypt, this should always be done, even if encryption for the
	-draft has not been enabled. In that case, the padlock should also be
	-crossed out. The actual icon and the coloring should vary according
	-to the degree to which the key is verified.
	+draft has not yet been enabled. In that case, the padlock should also
	+be crossed out. The actual icon and the coloring should vary
	+according to the degree to which the key is verified.

	It is important to distinguish between the different degrees of
	verification. The web of trust provides three verification levels: a
	key can either by fully verified, marginally verified or not verified.
	-(Note: GnuPG uses the term "trusted" here instead of "verified." To
	-reduce confusion in this document, we reserve the term trusted for
	-when a key is not just verified to be controlled by the stated entity,
	-but may act as an introducer.) And, the TOFU trust model provides
	-even finer grain verification levels. These distinctions are
	-important for security conscious users, and, as a rule of thumb,
	-marginally verified keys should not be shown as having the same
	-level of security as fully verified keys. Instead, data signed by
	-fully verified keys should be shown in, say, green, and data signed by
	-partially verified keys should be shown in, say, yellow. If it is
	-desirable that marginally verified keys have the same security level
	-as fully verified keys, then it is better to explicitly set the
	-~marginals-needed~ option in the user's ~gpg.conf~ file to ~1~.
	-
	-If the user hovers the mouse over the lock pad icon or clicks on it,
	+(Note: for historical reasons, GnuPG uses the term "trusted" here
	+instead of "verified." To reduce confusion in this document, we
	+reserve the term trusted for when a key is not just verified to be
	+controlled by the stated entity, but may act as an introducer. MUAs
	+should do the same.) And, the TOFU trust model provides even finer
	+grain verification levels. These distinctions are important for
	+security conscious users, and, as a rule of thumb, marginally verified
	+keys should not be shown as having the same level of security as
	+fully verified keys. Instead, fully verified keys should be shown in,
	+say, green, and partially verified keys should be shown in, say,
	+yellow. If it is desirable that marginally verified keys have the
	+same security level as fully verified keys, then it is better to
	+explicitly set the ~marginals-needed~ option in the user's ~gpg.conf~
	+file to ~1~.
	+
	+If the TOFU trust model is enabled, the number of days on which a
	+message has been encrypted to the key plus the number of days on which
	+a message signed by the key has been verified should be shown next to
	+the icon. This can be shown in a small bubble subscripting the icon,
	+which is similar to what Twitter does for showing counts. For large
	+numbers, it is reasonable to show approximate numbers (e.g., rounding
	+~1132~ to ~1.1k~).
	+
	+Showing these statistics is important to help users to detect mimicry
	+attacks, which are often employed by phishers. For instance, if a
	+bank normally signs their emails, then users hopefully become used to
	+seeing the count slowly increase. Then, if they get an email that
	+appears to be from their bank, but the count is ~0~, they will
	+hopefully become suspicious.
	+
	+If the user hovers the mouse over the padlock icon or clicks on it,
	the MUA should show a short, tweet-length message explaining why the
	key is considered verified (or not). If the key is not fully
	verified, an option to start the key verification wizard should be
	-provided. If there is no key associated with the email address, there
	-should be an option to start the key discovery wizard.
	+provided. If there is a TOFU conflict, there should be an option to
	+start the TOFU conflict resolution wizard. And, if there is no key
	+associated with the email address, there should be an option to start
	+the key discovery wizard. If the TOFU trust model is enabled and
	+there is a conflict, then the menu should include an option to start
	+the conflict resolution wizard.

	**** BCC Recipients

	When sending a mail, if there are any ~bcc~ recipients, the MUA should
	create a separate mail for each ~bcc~ recipient, and one for the rest.
	-Although it is possible to hide a recipient's key\nbsp{}ID in a
	-message by using a speculative key\nbsp{}ID (e.g., using ~gpg~ 's
	-~--throw-keyids~ option), this still reveals to the recipients that
	-the message was probably encrypted to other people. Using separate
	-emails avoids this leak.
	+This avoids having the OpenPGP implementation leak the ~bcc~
	+recipients to the other recipients. Although it is possible to hide a
	+recipient's key\nbsp{}ID in a message by using a speculative
	+key\nbsp{}ID (e.g., using ~gpg~ 's ~--throw-keyids~ option), this
	+still reveals to the recipients that the message was probably
	+encrypted to other people. Using separate emails avoids this leak.

	**** Saving Drafts

	-In general, if a draft---whether it has been marked to be encrypted
	+In general, when a draft---whether it has been marked to be encrypted
	/or/ not---is saved on the IMAP server, it should be encrypted to the
	-user (and not to any recipients, who should only be able to decrypt a
	-final version).
	+user. It should not be encrypted to any recipients; they should only
	+be able to decrypt the final version.

	-It is important to even encrypt drafts that have not been marked for
	-encryption, because the user's intent is only fully clear once the
	-mail has been sent. It may be reasonable to relax this requirement in
	-cases where it is clear that the user is only using the encryption for
	-privacy purposes. But another way to avoid decrypting the drafts is
	-to /also/ save an unencrypted copy locally.
	+It is important to encrypt all drafts even if they that have not been
	+marked for encryption, because the user's intent is only known once
	+the mail has been sent. It may be reasonable to relax this
	+requirement in cases where it is clear that the user is only using the
	+encryption for privacy purposes. But a safer way to avoid decrypting
	+the drafts is to /also/ either save the session key or an unencrypted
	+copy locally.

	**** Sent Mails

	When sending a mail, it is important to also encrypt the mail to the
	user. Given the near universal prevalence of a sent folder in MUAs,
	most users clearly expect to occasionally be able to later read the
	mails that they send. This can be done using ~gpg~ 's ~encrypt-to~
	option, or, when encrypting an email, the sender can be specified
	explicitly.

	**** Attaching Keys

	-It can be useful to have an option to attach the user's public key to
	-an email. Receiving a key can be surprising to users who don't use or
	-know about GnuPG. But, if you are encrypting, or signing an email,
	-this is not a concern: you know the recipients' MUA understands
	-OpenPGP messages.
	+To make it easier for a recipient to reply to a message in an
	+encrypted manner, the MUA should provide an option to attach all
	+public keys she would need to do so.
	+
	+Receiving a key can be surprising to users who don't use or know about
	+GnuPG. But, if you are encrypting, or signing an email, this is not a
	+concern: you know the recipients' MUA understands OpenPGP messages.
	+As such, in these cases, the keys can be attached automatically.

	When attaching a key, it is reasonable to just include a minimal
	-version of the key. In particular, you don't need to include any
	-certifications, because once the recipient has your key, it is easy to
	+version of the key. In particular, it doesn't need to include any
	+certifications, because once the recipient has the key, it is easy to
	get the rest of the data from a key server. A minimal key can be
	created by providing the option ~--export-options export-minimal~ when
	exporting a key using ~gpg~.

	The user's key should also always be specified in the OpenPGP
	-header\nbsp{}cite:cite:smasher2014openpgp-mail-header. This is the
	-case whether the mail is encrypted or not.
	+header\nbsp{}cite:smasher2014openpgp-mail-header. This is the case
	+whether the mail is encrypted or not. This provides a strong hint to
	+recipients that the user can work with OpenPGP messages.

	*** Reading Mail

	When the user opens an email message, it is necessary to identify if
	the message is encrypted or signed and to act accordingly. This is
	relatively straightforward, but does require a robust MIME parser to
	be able to handle all email. In particular, emails that have been
	transformed during transport can be problematic. The more challenging
	issue is making sure the user understands whether a message has been
	transferred securely. Other issues include avoiding unnecessary
	passphrase prompts, and searching encrypted email.

	**** Verifying Messages

	When a user views an email, it is important to communicate whether the
	contents were transferred in a secure fashion. In web browsers, this
	type of information is usually shown using a small padlock icon in the
	address bar.

	Firefox, for instance, shows a green padlock if it transferred the
	website in an encrypted manner, /and/ it could authenticate the
	end-point. It uses a gray padlock with a yellow warning triangle if
	some---but not all---of the content was encrypted, and eavesdropping
	was possible, or if the website used a self-signed certificate. It
	uses a gray padlock with red strikethrough if a man-in-the-middle
	attack was possible. And it just shows a neutral, "more information"
	-icon if TLS was not used at all\nbsp{}cite:mozilla-padlock.
	-
	-There are two important issues with this scheme.
	+icon if TLS was not used at all\nbsp{}cite:mozilla-padlock. There are
	+two important issues with this scheme.

	The first issue is that this scheme conflates encryption and
	authentication. Although it might be reasonable to demand that
	websites that use authentication also use encryption to be considered
	secure---it simplifies user training, and doesn't impose a significant
	deployment cost---this argument doesn't apply in an email setting.
	Consider, for instance, a company that wants to sign all of its
	outgoing emails to help mitigate phishing. In this scenario,
	-encryption is more of a hinderance than a help: requiring encryption
	+encryption is more of a hindrance than a help: requiring encryption
	would mean that the company would have to somehow find the right
	encryption key for each of its correspondents. When only providing an
	authentication mechanism, not only are the customers' keys not
	required, the customers don't even need to have a key: they just need
	the ability to validate the signature.

	The second problem is that a TLS connection that can't be
	-authenticated shown to be worse than a connection that is completely
	-insecure. For instance, until the recent introduction of /Let's
	-Encrypt/, website operators who wanted to offer an encrypted
	+authenticated is shown to be worse than a connection that is
	+completely insecure. For instance, until the recent introduction of
	+/Let's Encrypt/, website operators who wanted to offer an encrypted
	connection to their website, but didn't want to pay for a certificate
	could use a self-signed certificate. Although data protected by such
	certificates is not secure in the sense that the end point can't be
	authenticated without user intervention, such certificates enable
	encryption, which does protect users from passive surveillance. In
	other words, self-signed certificates provide more protection than
	nothing at all, but websites that use self-signed certificates are
	shown as being less secure than sites that use no protection at all!

	Happily, at least the Chrome browser does not make this distinction.
	And, like Chrome, we strongly recommend that whatever mechanism is
	used to show that a mail can't be authenticated be used for /both/
	unsigned mails, and mails with a signature that can't be verified.
	Specifically, we recommend considering an unencrypted and unsigned
	email to be the baseline, and that an email is never displayed in such
	a way that the user would consider it to be less secure than the
	baseline, unless there is strong evidence of an attack.

	It is reasonable to show unverified messages, and unsigned messages in
	a neutral manner, and to show verified messages in a positive manner.
	However, it may also be reasonable to show unverified messages, and
	unsigned messages in a negative manner. This is how MS Outlook
	behaves when S/MIME is enabled. This has the added advantage that it
	may prompt the user to learn why the MUA showed the message as being
	unsafe.

	The first step to checking whether a message is authentic is to check
	whether the signing key is verified according to some trust model,
	e.g., the web of trust. When verifying an email, another step is
	-required: it is also necessary to make sure that the key is controlled
	-by the sender. This can be done by checking that the email address in
	+required: it is also necessary to make sure the key is controlled by
	+the sender. This can be done by checking that the email address in
	the email's ~From~ header actually appears in one of the key's
	verified user\nbsp{}IDs. This is necessary to prevent an attacker
	from reusing a message in a different context. For instance, assuming
	Romeo trusts his father, his father could write an email that appears
	-to come from Juliet, but sign it with his own key. If we don't check
	-that the ~From~ header and the signer field agree, then Romeo would
	-see that the message is verified.
	+to come from Juliet, but sign it with his own key. If the MUA doesn't
	+check that the ~From~ header and the signer field agree, then the MUA
	+would show Romeo that the message is verified.

	Just checking that the sender matches a verified user ID is not
	actually enough to prevent all replay and mimicry attacks. It is also
	necessary to make sure the embedded timestamp is similar to (i.e.,
	within a few hours of) the email's timestamp. If the timestamp in the
	email is years later than the one embedded in the signature, then the
	email may be part of an attempted replay attack. Similarly, it is
	possible to change the recipient. For instance, Juliet might send the
	following signed message to Paris: "Go away, I do not love you!" But,
	Paris, realizing that Romeo and Juliet are in love, and hoping to
	trick Romeo, might simply send a copy of the message to him with the
	-From header set to Juliet. These types of attacks can be mitigated by
	-also verifying the mail headers. The Memory Hole project was started
	-to do exactly this\nbsp{}cite:memory-hole. Unfortunately, the
	-standard was never fully developed. Nevertheless, there is enough
	-information to understand the intent, and several mail clients
	-including Enigmail and Mailpile implement it.
	+~From~ header set to Juliet. These types of attacks can be mitigated
	+by also verifying the mail headers. The Memory Hole project was
	+started to do exactly this\nbsp{}cite:memory-hole. Unfortunately, the
	+standard isn't finished, and work on it appears to have stalled.
	+Nevertheless, there is enough information to understand the intent,
	+and several mail clients including Enigmail and Mailpile implement it.

	Sometimes, a message may include multiple signatures. Any signatures
	-from keys that match the email address in the From header should be
	+from keys that match the email address in the ~From~ header should be
	used for verification purposes. Other keys may be listed when showing
	the verification details.

	+If the TOFU trust model is enabled, then the TOFU statistics should be
	+shown as in the encryption case.
	+
	**** Multi-part Emails

	Thanks to inline signatures, it is trivial to make a message that is
	only partially verifiable.

	For simplicity's sake---we don't want to confuse the user---it is
	tempting to treat such messages as insecure like web browsers do.
	However, some companies, and some mailing lists automatically append a
	footer to all messages. This modification would change a message that
	is otherwise completely verifiable to one that contains a part that
	isn't signed. Thus, messages coming from these sources would never
	show up as secure.

	A simple solution is to show each section individually. This can be
	done using a frame. The frame should be part of the MUAs chrome and
	not the message to avoid mimicry attacks. Further, each part should
	-have an icon, e.g., a padlock, that shows information about the part
	-(the degree to which it is verified, in which an unsigned message
	-appears to be verified, and how that was computed) and that, when
	-clicked, shows a menu allowing the user to find or verify the key, as
	-appropriate.
	+have an icon, e.g., a padlock, that shows information about the part's
	+verification (the degree to which it is verified, and the key's TOFU
	+statistics), and that, when clicked, shows a menu allowing the user to
	+get more information, and find the key if it is missing, verify the
	+key if it is present, or resolve a TOFU conflict, as appropriate.

	To further distinguish between verified and unverified parts, a
	special background can be used. Ideally, the background should be
	unique for each user to further frustrate any attempt at a mimicry
	attack.

	The issues raised so far are manageable. Unfortunately, MIME makes
	things much more complicated: MIME makes it easy to not only transfer
	multi-part documents, but to transfer rich content that logically
	consists of multiple MIME parts only some of which are signed.

	If a message includes at least one verified part, then the MUA should
	only show those parts that are verified, and warn the user that the
	message contained unverified content that is hidden. It is reasonable
	for the warning to include an option to show the unverified parts
	anyway. At that point the message should be displayed as insecure.

	This suggestion conflicts with our earlier suggestion of showing
	unsigned messages in the same way as unverified messages. The best
	suggestion we have is to show a warning along the lines of "this
	message is unverified, show anyway" for unsigned messages. But, since
	most users will primarily deal with unsigned mail, this warning will
	-very quickly get annoying, and lose its value. If multiple profiles
	-are supported, this option should probably be enabled for users who
	-have high security requirements.
	+very quickly get annoying, and lose its value. If security profiles
	+are supported, this option should only be enabled for users who have
	+high security requirements.

	**** Unencrypted Cache

	The OpenPGP email workflow assumes that messages are stored on an
	untrusted server, and thus continue to need protection even after the
	mail has been delivered. Supporting this type of workflow is one of
	the primary reasons that OpenPGP doesn't provide forward secrecy.
	There are two major consequences of this workflow.

	First, every time a message is accessed, it needs to be decrypted.
	This can lead to many passphrase prompts. These can be largely
	mitigated by increasing the amount of time ~gpg-agent~ caches
	passphrases, or by using a password manager. But, it is also annoying
	for smartcard users who need to basically always leave their smartcard
	-inserted, which effectively nullifies a nice property of smartcards:
	-you can observe operations, because they can only be done when the
	-card is inserted.
	+inserted, which effectively nullifies a nice security property of
	+smartcards: the user can observe operations, because they can only be
	+done when the card is inserted.

	Second, it is not possible to search encrypt mails. This is a major
	usability problem, particularly, when the subject line is also
	obscured, as it should be to avoid unnecessarily leaking the message's
	contents.

	Both of these issues can be largely mitigated by caching the
	unencrypted version of each message locally. This assumes, of course,
	that the local device is secure. At a minimum, the user should have
	the mail stored on an encrypted partition.

	*** Key Management

	There are three main aspects to key management: key discovery, key
	verification, and key organization.

	**** Key Discovery

	The first requirement for encrypting or verifying a message is having
	the appropriate key. There are several ways to find the right key.
	-Unfortunately, most provide little guarantee about the key's validity.
	-But some are less easy to corrupt than others making them appropriate
	-for opportunistic encryption.
	+Unfortunately, most of them make no guarantee that the key that is
	+returned is the correct key. But some are significantly more
	+difficult for an adversary to corrupt than others making them at least
	+appropriate for opportunistic encryption.

	***** Exchanging Fingerprints in Person

	The most secure way to find a person's key is to get it from that
	person directly. If a physical meeting is possible, this can be done
	-by exchanging fingerprints in person. Because exchanging business
	-cards is a common practice in the business world anyway, adding your
	-fingerprint your business card makes securely exchange fingerprints a
	-free byproduct of a well-established ritual.
	-
	-The only remaining difficulty is entering the fingerprint into the
	-system. The key discovery wizard can make this process easier by
	-suggesting possible matches based on what the user has entered so far.
	-(Possible matches can be found by querying a key server.)
	+by exchanging fingerprints in person. At least in the business world,
	+the cost of this exchange can be driven to zero: because exchanging
	+business cards is a common practice in this world, adding your
	+fingerprint to your business card makes securely exchange fingerprints
	+a free byproduct of a well-established ritual.
	+
	+Having a fingerprint on a business card is not quite enough to use it:
	+it still needs to be entered into the system. The key discovery
	+wizard can make this process easier by suggesting possible matches
	+based on what the user has entered so far. (Possible matches can be
	+found by querying a key server.)

	We recommend having the user enter at least 64-bits worth of the
	fingerprint before enabling auto completion to ensure that the user
	checked a minimal amount of the fingerprint. For instance, it is
	-possible to create a chosen 32-bit key\nbsp{}ID in such a few seconds
	-on modern desktops\nbsp{}cite:evil32.
	+possible to create a key with a specific 32-bit key\nbsp{}ID in just a
	+few /seconds/ on modern desktop computers\nbsp{}cite:evil32.

	If the email address is known (and it is probably reasonable to first
	-ask the user to specify whose fingerprint she wants to enter), and
	-there is at least one matching key, an alternative approach is to show
	-a series of buttons with fragments of the matching fingerprints, and
	-have the user select the matching fragments. This idea is illustrated
	-below:
	+ask the user to specify a contact if this is not clear from the
	+context), and there is at least one matching key, an alternative
	+approach is to show a series of buttons with fragments of the matching
	+fingerprints, and have the user select the matching fragments. This
	+idea is illustrated below:

	#+BEGIN_EXAMPLE
	[ 8F17 ] [ 5200 ] [ 18A3 ]
	[ 3DDA ] [ 6396 ] [ 8723 ]
	[ AACB ] [ 6388 ] [ 0BAD ]

	[ None of the above ]
	#+END_EXAMPLE

	The "none of the above" option is useful if the right key is not on
	the key servers, for whatever reason.

	A more user-friendly technique could use a webcam and OCR to read in
	the fingerprint. From an implementation perspective, this is more
	-demanding than reading a QR\nbsp{}code, for instance, but there are
	+demanding than scanning a QR\nbsp{}code, for instance, but there are
	many fewer people who add a QR code containing their fingerprint to
	-their business card than who add their fingerprint.
	+their business card than those who add their fingerprint. But,
	+providing an option to display a public key using a QR\nbsp{}code on
	+screen can be helpful: someone could scan it.

	***** Picking up the Phone

	Exchanging keys in person requires that people actually meet face to
	face. This is often not practical. The next best alternative is to
	-pick up the phone. This is appropriate for all but those people who
	-have the highest security concerns---those whose threat model includes
	-a real-time voice imitator. Although this has been technically
	-feasible for years. It requires precise timing that only a
	-nation-state adversary could afford.
	+pick up the phone. This approach is appropriate for all but those
	+people who have the highest security concerns---those whose threat
	+model includes a real-time voice imitator. Although this has been
	+technically feasible for years. It requires precise timing that only
	+a nation-state adversary could afford.

	Again, assuming the email address is known, the button grid can be
	used to facilitate transcription of the fingerprint.

	***** Searching a Website

	-Sometimes calling is not possible or not desirable. In this case, it
	+Calling someone is not always possible or desirable. In this case, it
	is sometimes possible to find the person's key on her website. The
	caveats are that even a relatively unsophisticated attacker can often
	-modify a website or spoof it, and because there hasn't traditionally
	-been a standard place to publish keys, the MUA can't actually help the
	-user find it.
	+own a website or spoof it, and because there hasn't traditionally been
	+a standard place to publish keys, the MUA can't actually help the user
	+find it.

	In 2016, the GnuPG project published a new key discovery protocol
	called the web key directory (WKD). WKD automates, and hardens this
	key discovery process. The basic idea is that to find
	-~romeo@posteo.de~ 's key, Juliet looks in a database managed by the
	-user's mail provider, which in this case is
	-~posteo.de~\nbsp{}cite:koch2017wkd. The protocol relies on the
	-security of TLS, and the mail provider. The mail provider can
	-currently be held in check by periodically auditing the database.
	-Eventually, one could add something like certificate
	-transparency\nbsp{}cite:rfc6962 to catch abuse or detect things like
	-national security letters. The reliance on TLS and its centralized
	-infrastructure goes against the philosophy of OpenPGP, but it is
	-acceptable for people whose threat model is limited to privacy
	-violations, and phishing excursions.
	+~romeo@posteo.de~ 's key, Juliet looks for the key associated with
	+~romeo@posteo.de~ in a database on ~posteo.de~\nbsp{}cite:koch2017wkd.
	+This protocol relies on the security of TLS, and the mail provider.
	+The mail provider can currently be held in check by periodically
	+auditing the database, e.g., periodically fetching your own key via
	+Tor and making sure it hasn't been replaced. Eventually, something
	+like certificate transparency\nbsp{}cite:rfc6962 could be added to
	+catch abuse or detect things like national security letters. The
	+reliance on TLS and its centralized infrastructure goes against the
	+philosophy of OpenPGP, but it is acceptable for people whose threat
	+model is limited to privacy violations, and phishing excursions.

	Currently, the only commercial mail provider that supports WKD is
	-Posteo.de, but, as of 2017, there are discussions underway with other
	+Posteo, but, as of 2017, there are discussions underway with other
	mail providers to add support for this feature.

	***** Searching Key Servers

	A seemingly convenient way to find someone's key is to search for it
	using that person's email address on a public key server.
	Unfortunately, this method has very bad security properties: anyone
	can upload a key to a key server with any user\nbsp{}ID. It is
	-trivial to forge a user\nbsp{}ID. But, even if you are only
	-interested in the privacy aspects of encryption, the key servers are a
	-bad place to look for keys. Because many people forget their
	+trivial to forge a user\nbsp{}ID. In fact, in 2014, all known keys
	+were cloned with identical short key\nbsp{}IDs. But, even if you are
	+only interested in the privacy aspects of encryption, the key servers
	+are a bad place to look for keys. Because many people forget their
	passphrase or forget to migrate their key to a new computer, the key
	servers are littered with seemingly valid keys that are practically
	unusable. Since someone searching the key servers doesn't know what
	key is correct, these people often get emails that they can't decrypt.
	This is annoying, and causes people to avoid encryption.
	Consequently, if a MUA decides to provide support for looking up keys
	-by the user\nbsp{}ID, we strongly advise adding a strong warning about
	-the possible problems. Further, if this must be used, it is better to
	-encrypt to all matching keys. When the recipient replies, it is then
	-possible to narrow down the selection---assuming there was no man in
	-the middle attack.
	+by their user\nbsp{}ID, we strongly advise adding a prominent warning
	+about the possible problems. Further, if this approach must be used,
	+it is better to encrypt to all matching keys. When the recipient
	+replies, it is then possible to narrow down the set of potential keys
	+based on the signature or the PK-ESK packets---assuming there was no
	+man in the middle attack.

	Note: these problems don't mean that key servers are completely
	useless. Far from it. The problem with key servers is that
	user\nbsp{}IDs are not authoritative. But, if you have already
	verified someone's key, then key servers are the perfect place to get
	any updates (e.g., new signatures, revocation certificates, etc.),
	-because the cryptography can be used to determine that this
	-information really belongs to the key in question.
	+because cryptography can be used to determine whether the information
	+really belongs to the key in question.

	***** Exploiting Context and Hints

	There are two main places where context can be used to discover
	potentially useful keys: a signed message indicates what key was used
	to sign it, and an encrypted message usually includes the
	key\nbsp{}IDs of the sender and other recipients in the PK-ESK
	-packets. Mails also sometimes include hints about the right keys to
	-use. For instance, some people attach their key to mails that they
	-send, or the keys of all recipients to introduce people in a
	-multi-party discussion. Also, pEp does this as a matter of course.
	-And, another hint can often be found a mail's headers: there is a
	-standard for specifying the sender's
	+packets. Emails also sometimes include hints about the right keys to
	+use. For instance, some people attach either their key to the emails
	+that they send (pEp does this by default), or the keys of all
	+recipients in order to make it easier for people to reply in a
	+multi-party discuss. Another hint can sometimes be found among a
	+mail's headers: the ~OpenPGP~ header allows the sender to advertise a
	key\nbsp{}cite:smasher2014openpgp-mail-header.

	-In theory, there is no reason to not important these keys. Simply
	-importing a key will cause it to be considered verified: whether a key
	-is considered to be verified, is determined exclusively by the trust
	-model, not whether it happens to be available locally. But, having
	-what is probably the right key available locally is useful for
	+In theory, there is no reason to not import these keys. Simply
	+importing a key will not cause it to be considered verified: whether a
	+key is considered to be verified, is determined exclusively by the
	+trust model, not whether it happens to be available locally. But,
	+having what is probably the right key available locally is useful for
	opportunistic encryption. And, used in conjunction with the TOFU
	trust model, it is even possible to bootstrap some trust over time.

	-Unfortunately, there are two potential issues to harvesting keys.
	+Unfortunately, in practice there are two important issues with
	+harvesting keys.

	The first issue is that automatically fetching keys via the network
	can be used as a back channel. A sophisticated attacker could create
	-a new key for each message. When a user fetches the key, then if the
	-attacker observes the traffic, she knows not only that the user opened
	-the message, but also the user's IP address. This attack can be
	-mitigated by routing traffic via Tor (to do this, Tor must be
	-installed and GnuPG configured to use it by adding ~use-tor~ to
	-~dirmngr.conf~). Using Tor not only hides the user's IP address, but
	-also requires the attacker to actually control the user's preferred
	-key servers to observe the fetch. This is only feasible by an
	-adversary with a lot of resources.
	-
	-The second issue is GnuPG doesn't handle very large key rings (those
	-with thousands of keys) very well. This manifests itself in two ways.
	-It shows up with the longer random access times: ~gpg~ does a linear
	-scan of the key ring the first time the key ring is accessed. Also,
	+a new key for each message. When a user fetches the key, the attacker
	+can potentially learn not only that the user opened the message, but
	+also the user's IP address. This attack can be mitigated by routing
	+this type of traffic via Tor (to do this, Tor must be installed and
	+GnuPG configured to use it by adding ~use-tor~ to ~dirmngr.conf~).
	+Using Tor not only hides the user's IP address, but also requires the
	+attacker to actually control the user's preferred key servers to
	+observe the fetch. This is only feasible by an adversary with a lot
	+of resources.
	+
	+The second issue is that GnuPG doesn't handle very large key rings
	+(those with thousands of keys) very well. This manifests itself in
	+two ways. It shows up as longer random access times: ~gpg~ does a
	+linear scan of the key ring the first time it is accessed. Also,
	GnuPG's trust calculations are done on demand when ~gpg~ starts.
	These calculations can take minutes on large key rings. And, they are
	done whenever a new key or signature is imported, or a key expires or
	-is revoked. The trust calculations can be deferred by setting
	+is revoked. When harvesting keys, this can happen very often.
	+Happily, the trust calculations can be deferred by setting
	~no-auto-check-trustdb~ in ~gpg.conf~ and then running ~gpg
	---check-trustdb~ periodically, e.g., from something like ~cron~.
	-However, the only long-term fix is to improve the way that keys are
	-stored on disk.
	+--check-trustdb~ periodically, e.g., from something like ~cron~. But
	+obviously, this means the trust model may not be completely up to
	+date. However, the only long-term fix is to improve the way that keys
	+are stored on disk.

	Note: ~gpg~ can automatically fetch keys needed for verifying
	signatures by setting the ~auto-key-retrieve~ option in ~gpg.conf~,
	and for encrypting messages by setting the ~auto-key-locate~ option.
	These options have the disadvantage that they can potentially block
	the ~gpg~ process for a relatively long time. Consequently, it is
	often more appropriate to attempt to fetch the key in the background.
	-IN the verification case, the message can be rerendered if the key
	+In the verification case, the message can be rerendered if the key
	becomes available. And, in the encryption case, a key should be
	located in the background when the recipient is added, not when the
	message is sent.

	-***** Ask Trusted Introducers
	+***** Taking Advantage of Trusted Introducers

	-When the user users the web of trust and sets someone as a trusted
	-introducer, the user is explicitly saying that she trusts that
	-person's verifications. In that case, it makes sense to proactively
	-fetch any keys that such users have signed; these keys are known to be
	-good.
	+Designating someone as a trusted introducer means that the user trusts
	+that person to correctly verify others. Since friends of friends are
	+likely to be friends as well, it makes sense to proactively fetch any
	+keys that trusted introducers have signed.

	-~gpg~ does not do this itself. But, the MUA could do this
	-periodically in the background.
	+~gpg~ does not do this itself. And, unfortunately, the key servers do
	+not provide a mechanism to find all keys signed by a particular key.
	+But, since verification is usually mutual, it is possible to
	+approximate this by fetching all keys that signed a trusted
	+introducer's key. The MUA can do this periodically in the background.

	**** Key Verification

	Key verification is essential to the security of the system. Although
	people who are primarily interested in preserving their privacy will
	not spend much time on this task, it is essential that the key
	verification support is robust for those who depend on it for its
	security properties.

	This requirement first means that it should be easy to start the key
	verification wizard in appropriate contexts. For instance, when the
	user adds a recipient to an email, as explained above, an icon should
	-be shown whether there is a key associated with the contact, and, if
	-so, the degree to which the key is considered verified. Clicking on
	-the icon should allow the user to verify the key.
	-
	-When the key verification wizard is started, it should not just simply
	-prompt the user to check the fingerprint, but actually guide the user
	-through the different ways to obtain a fingerprint. Specifically, it
	-should ask the user how she wants to confirm the key: using a business
	-card or other printout, or via phone. This approach educates the user
	-without being patronizing: the user learns how to verify a
	-fingerprint, and that it is not okay to just click verify without
	-actually verifying the key.
	+be displayed showing whether there is a key associated with the
	+contact, and, if so, the degree to which the key is considered
	+verified. Clicking on the icon should allow the user to verify the
	+key.
	+
	+When the key verification wizard is started, it should not just prompt
	+the user to check the fingerprint, but actually guide the user through
	+the different ways to obtain a fingerprint. For instance, the
	+following is a bad idea:
	+
	+#+BEGIN_EXAMPLE
	+Certify this key?
	+
	+ 8F17 7771 18A3 3DDA 9BA4 8E62 AACB 3243 6300 52D9
	+
	+ [ Ok ] [ Cancel ]
	+#+END_EXAMPLE
	+
	+Instead, the key verification wizard should ask the user how she wants
	+to confirm the key: using a business card or other printout, or via
	+phone. This approach educates the user without being patronizing: the
	+user learns how to verify a fingerprint, and that it is not okay to
	+just click verify without actually verifying the key.

	To prevent the user from simply clicking okay without checking the
	fingerprint, we recommend requiring that the user enter at least part
	of the fingerprint. This can be done by using the buttons with the
	fingerprint fragments, as described above.

	***** Ownertrust

	It is strongly recommended that an option to set a key's ~ownertrust~
	be well hidden relative to the key verification option. In fact, it
	should only be possible to set the ~ownertrust~ if the key in question
	is already fully verified (e.g., directly signed). Also, even though
	there are a few rare cases where it makes sense, it shouldn't be
	possible to set a key to be ultimately trusted if no secret key
	material is available.

	When the ~ownertrust~ option is shown, it must be well explained that
	-it is not only about trusting the person, but also how she verifies
	-keys. In practice, the latter is generally much more difficult for
	-people to judge. For instance, I might trust my best friend when he
	-introduces me to people in the physical world, but without
	+this option is not only about trusting the person, but also trusting
	+how she verifies keys. For instance, I might trust my best friend
	+when he introduces me to people in the physical world, but without
	understanding how he verifies keys (does he just click on yes to make
	the padlock green?), I probably should not set him as a trusted
	-introducer. Given how hard it is to even get users to exchange
	-fingerprints, it is unlikely that we will convince them to discuss
	-their security practices with each other.
	+introducer. In practice, the latter is generally much more difficult
	+for people to judge, because they don't understand the process very
	+well themselves, and, given how hard it is to get people to exchange
	+fingerprints, it is unlikely that we will ever convince them to
	+discuss their security practices.
	+
	+***** Publishing Signatures
	+
	+The key verification wizard should provide an option to publish the
	+signature. This should be accompanied by an explanation of what this
	+means and why this is useful (people who trust you won't need to
	+manually verify this fingerprint).
	+
	+It is also reasonable to provide an option to make a trusted
	+signature instead of a simple certification. Again, this requires an
	+explanation. This option should probably only be hidden unless expert
	+mode is enabled.
	+
	+**** TOFU Conflict Resolution
	+
	+Like the web of trust, TOFU is a trust model. The major difference
	+between the two is that the web of trust provides strong guarantees,
	+but requires a lot of upfront verification work whereas TOFU builds up
	+trust slowly over time and is only secure in an asymptotic sense, but
	+requires little user support. The ~tofu+pgp~ trust model should be
	+the default for users with low security requirements. For backwards
	+compatibility reasons, TOFU has not been made the default in GnuPG.
	+
	+Normally, the user only needs to interact with the TOFU trust model to
	+resolve conflicts---when multiple valid keys have the same email
	+address. A conflict is a strong sign that a man-in-the-middle attack
	+is underway. But, it can also just be because the user replaced a key
	+that she could no longer access or revoke. The only way to resolve
	+this is by asking the user to verify the key. (When creating a new
	+key, a conflict can be avoided by either promptly revoking the old one
	+or cross signing the two keys.)
	+
	+When the user starts the conflict resolution wizard, the wizard should
	+explain what a conflict is, show the conflicting keys and their
	+statistics, and explain how to resolve the problem (ideally, the user
	+should call the contact to verify the fingerprint). Because the user
	+might not be able to resolve the conflict immediately, it is better to
	+provide a resolve later option, which is the default, rather than have
	+the user simply accept the key without validating it.
	+
	+Note: just because a key has a lot of past usage does not mean that it
	+is the right key: the man in the middle might just have failed to
	+intercept the most recent message. Likewise, the new key is not
	+necessarily the right one: the man in the middle might just have
	+started the attack.

	**** Address Book Integration

	A key ring is effectively a backwards address book: instead of names
	-being the primary keys, and the OpenPGP key being associated with
	-people, a key ring reverses this. This unusual arrangement can cause
	-normal users significant confusion. As such, it is better to avoid
	-the key ring as much as possible, and instead directly integrate keys
	-into the user's address book.
	+being the primary keys, and OpenPGP keys being associated with
	+contacts, a key ring reverses this. This unusual arrangement can
	+cause novice users significant confusion. As such, it is better to
	+avoid the key ring as much as possible, and instead directly integrate
	+keys into the user's address book.

	If the address book supports identities with multiple email addresses,
	-then it should be possible to associate these with multiple keys.
	-This is a convenient grouping function. Similarly, it should be
	-possible to force messages sent to a particular person to be encrypted
	-to multiple keys. This is useful in the case where an email address
	-acts as an exploder.
	+then it should be possible to associate each email address with a
	+different key. Also, it should be possible to force messages sent to
	+a particular contact to be encrypted to multiple keys. This is useful
	+in the case where an email address acts as an exploder.
	+
	+It is also useful to keep track of users who appear to use GnuPG. A
	+recent encrypted or signed email is the best indicator, but the
	+presence of the OpenPGP mail header is also an excellent hint. The
	+presence of a key with the user's email address is, however, not
	+sufficient proof that the user can use GnuPG. This functionality can
	+also be exposed as an option: "always encrypt to this user."

	** Programming with GnuPG

	~--batch~, ~--status-fd~ and ~--command-fd~.

	Writing tests: use ~--faked-system-time~. (Talk about how it works.)
	~gpg-compose~ for creating test data.

	Talk about GPGME.

	** Misc.

	Topics are probably better integrated someplace else:

	- gpgv

	- ~/etc/skel/.gnupg~

	- keyring vs. keybox. Talk about kbxutil.

	- What's a keygrip.

	- More tools: encrypted mailing lists (schleuder), form encryption
	(Kuvert),

	bibliographystyle:unsrt
	bibliography:bib.bib

	# LocalWords: cryptographic decrypted decrypts reauthenticate hoc
	# LocalWords: decrypting constrainted TripleDES OpenPGP's subkey
	# LocalWords: packetized subkeys unforgable subpackets unexportable
	# LocalWords: introducer Introducers subpacket OpenPGP SED AES WKD
	# LocalWords: cryptosystem ESK auditable reencrypted decrypt HTTP's
	# LocalWords: unbuffered apriori plaintext encypted data's OPS's
	# LocalWords: revoker verifications unhashed introducers natively
	# LocalWords: strikethrough MUA ownertrust Mailpile cron untrusted
	# LocalWords: KMail Enigmail MUAs unencrypted discoverability TLS
	# LocalWords: Snowden Autocrypt rerender cryptographically exploder
	-# LocalWords: opsec WKS BCC IMAP rerendered
	+# LocalWords: opsec WKS BCC IMAP rerendered nbsp GPGTools gpgol RSA
	+# LocalWords: Schneier gpg subscripting bcc introducer's

File Metadata

Mime Type: text/x-diff
Expires: Fri, Feb 6, 8:11 AM (1 h, 37 m)
Storage Engine: local-disk
Storage Format: Raw Data
Storage Handle: 19/b8/a6a60289feebd6aca52342e08222

No OneTemporaryActions

View Options

File Metadata

Event Timeline

No OneTemporary
Actions