More thoughts on working with GnuPG
Smári's blog post on working with GnuPG got a fair bit of attention and continues to be referenced in various posts and discussions online.
Responses have varied, from vehement agreement, to a friendly and thoughtful response from Werner himself, to random folks on the Internet people calling us incompetent for using the GnuPG command-line tool directly instead of some library or other.
I would like to address that last group specifically:
If you think a command-line interface is not an API, then you are ignoring the millions of lines of shell scripts that keep the Internet running. You are wrong. Please consult your nearest sysadmin for an attitude readjustment.
Process separation is one of the cornerstones of good security design. Heartbleed and many other vulnerabilities are direct results of crypto and keys living the same process space as protocol and application logic. We initially chose to invoke GnuPG (and OpenSSL as well) as external tools for such security reasons, even though we knew it would cost us some performance.
We looked at the libraries, encountered an apalling lack of documentation and hoped the command-line interface would be nicer to work with. Learning one API beats having to learn two, and there was no way we could do this without the GnuPG CLI - if only for debugging and tech support reasons.
Now that I've gotten that off my chest... let's move on to some more productive discussions!
Utility or Application?
Although I appreciated Werner's response to our post, I am afraid he missed the forest for the trees. Probably because Smári wrote a lot about some interesting, funky trees... :-)
Fundamentally, I think Mailpile's difficulties stem from the fact that GnuPG was not written with automation in mind. The tool is not written as a unix utility, it is very much written as a user facing application, a tool for humans.
As such, it is a fantastic piece of engineering. If you understand how public/private key encryption works and know what you want to do, odds are GnuPG will help you get the job done. GnuPG's security track record is solid and it is a tool people can rely on, as long as they take the time to learn the basics and get used to the CLI.
However, as a result of this design focus, many of GnuPG's automation interfaces appear to be an afterthought. They are incomplete and often difficult to work with. Consider that "commands to create and sign keys without any user interaction" were brand new in GnuPG 2.1 last November. GnuPG 2.1 is the development version of GnuPG and is not shipped by any of the mainstream Linux distributions yet. I do applaud the fact that these features are being added, but the fact that it happened 3 months ago only supports my hypothesis: automation has to a great extent been an afterthought to the GnuPG project.
This has gotten worse over time, not better. In GnuPG 2.0, use of
pinentry became mandatory and the assumption that the
tool is running in a desktop environment under human oversight pervades.
This is why Linux distros still keep shipping GnuPG 1.4.x - they rely on
automated PGP processing for package verification and other things which
GnuPG 2.0 just isn't suitable for.
I worry that GnuPG 2.1 takes this even further still, moving key
managment operations into the
agent. Also, as far
as I can tell, it doesn't matter whether we use the
gpgme library or
the command-line tool, those assumptions and behaviors are shared
Much of this is understandable, given GnuPG's historic roots and the fact that it inherits many things from it's predecessor, PGP, which itself predated modern graphical desktops and web apps. It is also understandable, given the culture around PGP, where keys and passphrases are magical holy talismans, so special that you need to constantly ask the user for permission before invoking them. GnuPG enforces this culture by preventing apps from interfering with the flow of this critical information, refusing to work without trusted pinentry.
However, while GnuPG insists on controlling this part of the user interface and treats automation as a 2nd class citizen, a side effect is that the project actively thwarts efforts of others that want or need alternate user interfaces.
Mailpile is one such project.
Mailpile's pinentry woes
One of the cool things about Mailpile, is it functions not only in a desktop environment, but also as a background process (receiving, decrypting and indexing mail), and also as a remote web server which allows the user to leave their mail and keys on a secure box in their home, which they then access remotely over the network.
This flexibility is key to our overall goals of providing people with something that can match "cloud based" e-mail providers on features, without compromising privacy and security.
Unfortunately, the entire pinentry concept, where
gpg (via the
gnupg-agent) displays a window prompting the user to enter a
passphrase, is completely and utterly incompatible with this. Popping up
a dialog on the user's Raspberry Pi, when the owner tries to read
encrypted mail from halfway across the world using his Android phone's
web browser, is never going to work.
And even if it somehow did work, using a side-channel to authenticate has very serious problems of its own.
Consider the situation where a user wants to access Mailpile both locally, and remotely: on the desktop, the "just use pinentry" dogma tells us we should just trust GnuPG there (and GnuPG 2.x enforces this policy).
Well, what then happens when Mailpile is accessed remotely?
The benign failure mode is the user can't read encrypted mail because they can't see the pinentry dialog. This makes the app unusable, but the bug is still "benign" because it doesn't leak confidential data.
There is another catastrophic (and real) failure mode: If the legitimate
user is logged on at the desktop and an attacker can access Mailpile's
web interface at the same time, then the attacker will be able to read
encrypted mail (or compose fraudulent signed mail) because the
gpg-agent has keys cached and Mailpile has no way to communicate to
GnuPG that the remote access is somehow different from what happens
This is really, really bad.
This is why Mailpile only supports GnuPG 1.4.x at the moment: we haven´t figured out how to make GnuPG 2.x operate in a secure fashion, since it won't let us disable the agent.
This whole issue has us scratching our heads and wondering what to do. We've gone from frustration to righteous anger, to just being rather worried. If the GnuPG project is moving aggressively in a direction which is fundamentally incompatible with projects like ours, what should we do? Drop PGP and try to implement S/MIME? Switch to something interesting but unproven, like reop?
I just don't know.
I could come up with some clever hacks to work around GnuPG's core design; I could implement my own agent, I could try to cripple GnuPG with weird environment variables or a custom $GNUPGHOME and config, or I could even fork the project - but those all feel like last resorts.
I am currently revisiting whether I can get
gpgme-based code to avoid
the problems caused by the agent and side-channel auth. I am not too
optimistic. If that fails, then it's last resort time, or time to drop
Update, 2015-06-22: After discussing in person with Werner of GnuPG, the current strategy is indeed one of the "last resorts". Mailpile will use a custom configuration file and avoid GnuPG entirely for certain local encryption and authentication operations. PGP, via GnuPG, will still be supported for incoming and outgoing e-mail.
Is this fixable?
However, I have yet to hear any of the GnuPG developers acknowledge that this is even a concern, which is why I have taken the time to write this post. I would like to humbly request that Werner and GnuPG consider these issues, now that they finally have a budget. (Congrats, by the way, you deserve it!)
I really do think this can be fixed, if the GnuPG project agrees it is worth doing. Some specific ideas that I think could help:
Clearly separate the user interface aspects of GnuPG from the crypto and keychain management, so projects can use one without the other. Do not assume all apps interacting with a particular keychain want the same UI for authentication.
Treat the command-line interface of
gpgas if it were an API; keep it stable and machine friendly. There should be command-line alternatives to all of the interactive dialogs.
gpgme-tool(which implements an API very similar to the JSON API requested by Smári in his post) a mandatory 1st class citizen of all GnuPG bundles (instead of doing this). We had no idea this tool existed, and I still can't find it for my distro (Ubuntu 14.04). Consider rolling this functionality into the
gpgbinary itself to facilitate scripting and automation.
There are probably more, but those are my top three!
Thanks for reading, I look forward to further discussions (most likley on the GnuPG Users mailing list).