Guix will hold its first online conference November 22nd. Propose a talk before November 6th. Learn more!

Securing updates

Software deployment tools like Guix are in a key position when it comes to securing the “software supply chain”—taking source code fresh from repositories and providing users with ready-to-use binaries. We have been paying attention to several aspects of this problem in Guix: authentication of pre-built binaries, reproducible builds, bootstrapping, and security updates.

A couple of weeks ago, we addressed the elephant in the room: authentication of Guix code itself by guix pull, the tool that updates Guix and its package collection. This article looks at what we set out to address, how we achieved it, and how it compares to existing work in this area.

What updates should be protected against

The problem of securing distro updates is often viewed through the lens of binary distributions such as Debian, where the main asset to be protected are binaries themselves. The functional deployment model that Guix and Nix implement is very different: conceptually, Guix is a source distribution, like Gentoo if you will.

Pre-built binaries are of course available and very useful, but they’re optional; we call them substitutes because they’re just that: substitutes for local builds. When you do choose to accept substitutes, they must be signed by one of the keys you authorized (this has been the case since version 0.6 in 2014).

Guix consists of source code for the tools as well as all the package definitions—the distro. When users run guix pull, what happens behind the scene is equivalent to git clone or git pull. There are many ways this can go wrong. An attacker can trick the user into pulling code from an alternate repository that contains malicious code or definitions for backdoored packages. This is made more difficult by the fact that code is fetched over HTTPS from Savannah by default. If Savannah is compromised (as happened in 2010), an attacker can push code to the Guix repository, which everyone would pull. The change might even go unnoticed and remain in the repository forever. An attacker with access to Savannah can also reset the main branch to an earlier revision, leading users to install outdated software with known vulnerabilities—a downgrade attack. These are the kind of attacks we want to protect against.

Authenticating Git checkouts

If we take a step back, the problem we’re trying to solve is not specific to Guix and to software deployment tools: it’s about authenticating Git checkouts. By that, we mean that when guix pull obtains code from Git, it should be able to tell that all the commits it fetched were pushed by authorized developers of the project. We’re really looking at individual commits, not tags, because users can choose to pull arbitrary points in the commit history of Guix and third-party channels.

Checkout authentication requires cryptographically signed commits. By signing a commit, a Guix developer asserts that they are the one who made the commit; they may be its author, or they may be the person who applied somebody else’s changes after review. It also requires a notion of authorization: we don’t simply want commits to have a valid signature, we want them to be signed by an authorized key. The set of authorized keys changes over time as people join and leave the project.

To implement that, we came up with the following mechanism and rule:

  1. The repository contains a .guix-authorizations file that lists the OpenPGP key fingerprints of authorized committers.
  2. A commit is considered authentic if and only if it is signed by one of the keys listed in the .guix-authorizations file of each of its parents. This is the authorization invariant.

(Remember that Git commits form a directed acyclic graph (DAG) where each commit can have zero or more parents; merge commits have two parent commits, for instance. Do not miss Git for Computer Scientists for a pedagogical overview!)

Let’s take an example to illustrate. In the figure below, each box is a commit, and each arrow is a parent relationship:

Example commit graph.

This figure shows two lines of development: the orange line may be the main development branch, while the purple line may correspond to a feature branch that was eventually merged in commit F. F is a merge commit, so it has two parents: D and E.

Labels next to boxes show who’s in .guix-authorizations: for commit A, only Alice is an authorized committer, and for all the other commits, both Bob and Alice are authorized committers. For each commit, we see that the authorization invariant holds; for example:

The authorization invariant has the nice property that it’s simple to state, and it’s simple to check and enforce. This is what guix pull implements. If your current Guix, as returned by guix describe, is at commit A and you want to pull to commit F, guix pull traverses all these commits and checks the authorization invariant.

Once a commit has been authenticated, all the commits in its transitive closure are known to be already authenticated. guix pull keeps a local cache of the commits it has previously authenticated, which allows it to traverse only new commits. For instance, if you’re at commit F and later update to a descendant of F, authentication starts at F.

Since .guix-authorizations is a regular file under version control, granting or revoking commit authorization does not require special support. In the example above, commit B is an authorized commit by Alice that adds Bob’s key to .guix-authorizations. Revocation is similar: any authorized committer can remove entries from .guix-authorizations. Key rotation can be handled similarly: a committer can remove their former key and add their new key in a single commit, signed by the former key.

The authorization invariant satisfies our needs for Guix. It has one downside: it prevents pull-request-style workflows. Indeed, merging the branch of a contributor not listed in .guix-authorizations would break the authorization invariant. It’s a good tradeoff for Guix because our workflow relies on patches carved into stone tablets (patch tracker), but it’s not suitable for every project out there.

Bootstrapping

The attentive reader may have noticed that something’s missing from the explanation above: what do we do about commit A in the example above? In other words, which commit do we pick as the first one where we can start verifying the authorization invariant?

We solve this bootstrapping issue by defining channel introductions. Previously, one would identify a channel simply by its URL. Now, when introducing a channel to users, one needs to provide an additional piece of information: the first commit where the authorization invariant holds, and the fingerprint of the OpenPGP key used to sign that commit (it’s not strictly necessary but provides an additional check). Consider this commit graph:

Example commit graph with introduction.

On this figure, B is the introduction commit. Its ancestors, such as A are considered authentic. To authenticate, C, D, E, and F, we check the authorization invariant.

As always when it comes to establishing trust, distributing channel introductions is very sensitive. The introduction of the official guix channel is built into Guix. Users obtain it when they install Guix the first time; hopefully they verify the signature on the Guix tarball or ISO image, as noted in the installation instructions, which reduces chances of getting the “wrong” Guix, but it is still very much trust-on-first-use (TOFU).

For signed third-party channels, users have to provide the channel’s introduction in their channels.scm file, like so:

(channel
  (name 'my-channel)
  (url "https://example.org/my-channel.git")
  (introduction
   (make-channel-introduction
    "6f0d8cc0d88abb59c324b2990bfee2876016bb86"
    (openpgp-fingerprint
     "CABB A931 C0FF EEC6 900D  0CFB 090B 1199 3D9A EBB5"))))

The guix describe command now prints the introduction if there’s one. That way, one can share their channel configuration, including introductions, without having to be an expert.

Channel introductions also solve another problem: forks. Respecting the authorization invariant “forever” would effectively prevent “unauthorized” forks—forks made by someone who’s not in .guix-authorizations. Someone publishing a fork simply needs to emit a new introduction for their fork, pointing to a different starting commit.

Last, channel introductions give a point of reference: if an attacker manipulates branch heads on Savannah to have them point to unrelated commits (such as commits on an orphan branch that do not share any history with the “official” branches), authentication will necessarily fail as it stumbles upon the first unauthorized commit made by the attacker. In the figure above, the red branch with commits G and H cannot be authenticated because it starts from A, which lacks .guix-authorizations and thus fails the authorization invariant.

That’s all for authentication! I’m glad you read this far. At this point you can take a break or continue with the next section on how guix pull prevents downgrade attacks.

Downgrade attacks

An important threat for software deployment tools is downgrade or roll-back attacks. The attack consists in tricking users into installing older, known-vulnerable software packages, which in turn may offer new ways to break into their system. This is not strictly related to the authentication issue we’ve been discussing, except that it’s another important issue in this area that we took the opportunity to address.

Guix saves provenance info for itself: guix describe prints that information, essentially the Git commits of the channels used during git pull:

$ guix describe
Generation 149  Jun 17 2020 20:00:14    (current)
  guix 8b1f7c0
    repository URL: https://git.savannah.gnu.org/git/guix.git
    branch: master
    commit: 8b1f7c03d239ca703b56f2a6e5f228c79bc1857e

Thus, guix pull, once it has retrieved the latest commit of the selected branch, can verify that it is doing a fast-forward update in Git parlance—just like git pull does, but compared to the previously-deployed Guix. A fast-forward update is when the new commit is a descendant of the current commit. Going back to the figure above, going from commit A to commit F is a fast-forward update, but going from F to A or from D to E is not.

Not doing a fast-forward update would mean that the user is deploying an older version of the Guix currently used, or deploying an unrelated version from another branch. In both cases, the user is at risk of ending up installing older, vulnerable software.

By default guix pull now errors out on non-fast-forward updates, thereby protecting from roll-backs. Users who understand the risks can override that by passing --allow-downgrades.

Authentication and roll-back prevention allow users to safely refer to mirrors of the Git repository. If git.savannah.gnu.org is down, one can still update by fetching from a mirror, for instance with:

guix pull --url=https://github.com/guix-mirror/guix

If the repository at this URL is behind what the user already deployed, or if it’s not a genuine mirror, guix pull will abort. In other cases, it will proceed.

Unfortunately, there is no way to answer the general question “is X the latest commit of branch B ?”. Rollback detection prevents just that, rollbacks, but there’s no mechanism in place to tell whether a given mirror is stale. To mitigate that, channel authors can specify, in the repository, the channel’s primary URL. This piece of information lives in the .guix-channel file, in the repository, so it’s authenticated. guix pull uses it to print a warning when the user pulls from a mirror:

$ guix pull --url=https://github.com/guix-mirror/guix
Updating channel 'guix' from Git repository at 'https://github.com/guix-mirror/guix'...
Authenticating channel 'guix', commits 9edb3f6 to 3e51f9e (44 new commits)...
guix pull: warning: pulled channel 'guix' from a mirror of https://git.savannah.gnu.org/git/guix.git, which might be stale
Building from this channel:
  guix      https://github.com/guix-mirror/guix 3e51f9e
…

So far we talked about mechanics in a rather abstract way. That might satisfy the graph theorist or the Git geek in you, but if you’re up for a quick tour of the implementation, the next section is for you!

A long process

We’re kinda celebrating these days, but the initial bug report was opened… in 2016. One of the reasons was that we were hoping the general problem was solved already and we’d “just” have to adapt what others had done. As for the actual design: you would think it can be implemented in ten lines of shell script invoking gpgv and git. Perhaps that’s a possibility, but the resulting performance would be problematic—keep in mind that users may routinely have to authenticate hundreds of commits. So we took a long road, but the end result is worth it. Let’s recap.

Back in April 2016, committers started signing commits, with a server-side hook prohibiting unsigned commits. In July 2016, we had proof-of-concept libgit2 bindings with the primitives needed to verify signatures on commits, passing them to gpgv; later Guile-Git was born, providing good coverage of the libgit2 interface. Then there was a two-year hiatus during which no code was produced in that area.

Everything went faster starting from December 2019. Progress was incremental and may have been hard to follow, even for die-hard Guix hackers, so here are the major milestones:

Whether you’re a channel author or a user, the feature is now fully documented in the manual, and we’d love to get your feedback!

SHA-1

We can’t really discuss Git commit signing without mentioning SHA-1. The venerable crytographic hash function is approaching end of life, as evidenced by recent breakthroughs. Signing a Git commit boils down to signing a SHA-1 hash, because all objects in the Git store are identified by their SHA-1 hash.

Git now relies on a collision attack detection library to mitigate practical attacks. Furthermore, the Git project is planning a hash function transition to address the problem.

Some projects such as Bitcoin Core choose to not rely on SHA-1 at all. Instead, for the commits they sign, they include in the commit log the SHA512 hash of the tree, which the verification scripts check.

Computing a tree hash for each commit in Guix would probably be prohibitively costly. For now, for lack of a better solution, we rely on Git’s collision attack detection and look forward to a hash function transition.

As for SHA-1 in an OpenPGP context: our authentication code rejects SHA-1 OpenPGP signatures, as recommended.

Related work

A lot of work has gone into securing the software supply chain, often in the context of binary distros, sometimes in a more general context; more recent work also looks into Git authentication and related issues. This section attempts to summarize how Guix relates to similar work that we’re aware of in these two areas. More detailed discussions can be found in the issue tracker.

The Update Framework (TUF) is a reference for secure update systems, with a well-structured spec and a number of implementations. TUF is a great source of inspiration to think about this problem space. Many of its goals are shared by Guix. Not all the attacks it aims to protect against (Section 1.5.2 of the spec) are addressed by what’s presented in this post: indefinite freeze attacks, where updates never become available, are not addressed per se (though easily observable), and slow retrieval attacks aren’t addressed either. The notion of role is also something currently missing from the Guix authentication model, where any authorized committer can touch any files, though the model and .guix-authorizations format leave room for such an extension.

However, both in its goals and system descriptions, TUF is biased towards systems that distribute binaries as plain files with associated meta-data. That creates a fundamental impedance mismatch. As an example, attacks such as fast-forward attacks or mix-and-match attacks don’t apply in the context of Guix; likewise, the repository depicted in Section 3 of the spec has little in common with a Git repository.

Developers of OPAM, the OCaml package manager, adapted TUF for use with their Git-based package repository, later updated to write Conex, a separate tool to authenticate OPAM repositories. OPAM is interesting because like Guix it’s a source distro and its package repository is a Git repository containing “build recipe”. To date, it appears that opam update itself does not authenticate repositories though; it’s up to users or developer to run Conex.

Another very insightful piece of work is the 2016 paper On omitting commits and committing omissions. The paper focuses on the impact of malicious modifications to Git repository meta-data. An attacker with access to the repository can modify, for instance, branch references, to cause a rollback attack or a “teleport” attack, causing users to pull an older commit or an unrelated commit. As written above, guix pull would detect such attacks. However, guix pull would fail to detect cases where metadata modification does not yield a rollback or teleport, yet gives users a different view than the intended one—for instance, a user is directed to an authentic but different branch rather than the intended one. The “secure push” operation and the associated reference state log (RSL) the authors propose would be an improvement.

Wrap-up and outlook

Guix now has a mechanism that allows it to authenticate updates. If you’ve run guix pull recently, perhaps you’ve noticed additional output and a progress bar as new commits are being authenticated. Apart from that, the switch has been completely transparent. The authentication mechanism is built around the commit graph of Git; in fact, it’s a mechanism to authenticate Git checkouts and in that sense it is not tied to Guix and its application domain. It is available not only for the main guix channel, but also for third-party channels.

To bootstrap trust, we added the notion of channel introductions. These are now visible in the user interface, in particular in the output of guix describe and in the configuration file of guix pull and guix time-machine. While channel configuration remains a few lines of code that users typically paste, this extra bit of configuration might be intimidating. It certainly gives an incentive to provide a command-line interface to manage the user’s list of channels: guix channel add, etc.

The solution here is built around the assumption that Guix is fundamentally a source-based distribution, and is thus completely orthogonal to the public key infrastructure (PKI) Guix uses for the signature of substitutes. Yet, the substitute PKI could probably benefit from the fact that we now have a secure update mechanism for the Guix source code: since guix pull can securely retrieve a new substitute signing key, perhaps it could somehow handle substitute signing key revocation and delegation automatically? Related to that, channels could perhaps advertise a substitute URL and its signing key, possibly allowing users to register those when they first pull from the channel. All this requires more thought, but it looks like there are new opportunities here.

Until then, if you’re a user or a channel author, we’d love to hear from you! We’ve already gotten feedback that these new mechanisms broke someone’s workflow; hopefully it didn’t break yours, but either way your input is important in improving the system. If you’re into security and think this design is terrible or awesome, please do provide feedback.

It’s a long and article describing a long ride on a path we discovered as we went, and it felt like an important milestone to share!

Acknowledgments

Thanks to everyone who provided feedback, ideas, or carried out code review during this long process, notably (in no particular order): Christopher Lemmer Webber, Leo Famulari, David Thompson, Mike Gerwitz, Ricardo Wurmus, Werner Koch, Justus Winter, Vagrant Cascadian, Maxim Cournoyer, Simon Tournier, John Soo, and Jakub Kądziołka. Thanks also to janneke, Ricardo, Marius, and Simon for reviewing an earlier draft of this post.

About GNU Guix

GNU Guix is a transactional package manager and an advanced distribution of the GNU system that respects user freedom. Guix can be used on top of any system running the kernel Linux, or it can be used as a standalone operating system distribution for i686, x86_64, ARMv7, and AArch64 machines.

In addition to standard package management features, Guix supports transactional upgrades and roll-backs, unprivileged package management, per-user profiles, and garbage collection. When used as a standalone GNU/Linux distribution, Guix offers a declarative, stateless approach to operating system configuration management. Guix is highly customizable and hackable through Guile programming interfaces and extensions to the Scheme language.