Improving Hackage security

Thursday, 16 April 2015, by Duncan Coutts.
Filed under industrial-haskell-group, cabal, community.

The members of the Industrial Haskell Group are funding work to improve the security of packages downloaded from Hackage. The IHG members agreed to fund this task some time ago and my colleague Austin and I have been working on the design and implementation.

In this post I will explain the problem and the design of our solution.

TL;DR

We’re implementing a system to significantly improve Hackage security. It’s based on a sensible design (The Update Framework) by proper crypto experts. The basic system is fully automatic and covers all packages on Hackage. A proposed extension would give further security improvements for individual packages at the cost of a modest effort from package authors.

It will also allow the secure use of untrusted public Hackage mirrors, which is the simplest route to better Hackage reliability. As a bonus we’re including incremental index downloads to reduce cabal update wait times. And it’s all fully backwards compatible.

Goals

The general aim is to make a significant incremental improvement to Hackage security, given limited resources to spend on the project. So that means covering as many users, packages and likely threats as we can.

We have had to think carefully about the implementation effort and what improvements will give the biggest security benefits. Our general approach has been to go for the biggest improvements first, but to leave the door open to further tightening later.

Crypto-humility and initial thoughts

Both Austin and I have some experience with cryptography and computer security. Austin used to work for a computer security company, and many years ago I worked at an anti-virus company on the problem of securing the software update process. So we know the general options. Importantly, we also know enough to know that we don’t know everything. Hopefully the combination will be enough to avoid security disasters like the recent Docker image download vulnerability.

The basic options are:

Our initial thoughts were that having the server sign the index gives a bigger bang for the buck than a scheme with author signing. This is because with the server doing the signing we can cover all packages and not require any knowledge or effort by either the package authors or the developers downloading the packages. With author signing we fear we would not be able to cover the majority of packages, at least not in the near term. Without covering all the packages we cannot have untrusted Hackage mirrors. Our rough estimate was also that index signing would be less effort to implement. Of course we’re well aware that index signing provides shallower security than the end-to-end verification with author signing.

So while we started with a general idea of the kind of solutions that would fit with Hackage and our budget, we looked around for designs and analysis from better experts than ourselves.

Austin pointed out some prior art called The Update Framework (abbreviated TUF). TUF is an architecture for securing software update systems, though there is also a reference implementation in Python. It has been designed by a number of academics, based on a few published papers and practical experience with the Tor Project’s update system. There is ongoing work to use TUF to secure Python and Ruby’s package distribution systems (both of which are fairly similar to Hackage).

What is good about TUF is that it considers the problem as a whole, with a wide range of threats. It has what security researchers call a “threat model”, which is needed to be able to discuss with clarity what kinds of threats a system can protect against. As an architecture it is also flexible enough to cover a fairly wide range of update systems: one can use subsets and still be clear about what protections one has.

The elements of TUF are still based on the basic ideas of end-to-end signing and index signing, but it is put together in a comprehensive way. There are several attacks that TUF considers and protects against that we did not consider with our initial thoughts on index signing.

Indeed even if one goes for a system based only on end-to-end author-based signing (e.g. individual GPG-signed packages) then there are still several possible attacks. In retrospect this is clear, but it was a bit of a surprise since packages individually GPG-signed by their authors is often considered as the gold standard.

The Update Framework

TUF is designed to protect against adversaries that can interfere with network traffic between you and the repository server, for example someone operating a wifi network you’re connected to, or in control of a proxy or operating a mirror of the repository. To a degree it is also designed to protect against adversaries than can compromise a repository server itself.

In a full implementation of the architecture, TUF subsumes both server-based index signing and end-to-end verification with author signing.

It relies on a number of different cryptographic keys with different roles. There are one or more root keys and a few subordinate keys for particular purposes. It does not rely on an external PKI, like the public certificate authority system that SSL/TLS uses.

The key idea with having multiple keys with different roles is that one can design a system in such a way that compromising one key is not enough to break the whole system. This relies on different keys living in different places so that compromising one doesn’t automatically imply compromising the others. Some keys have to live on public facing servers and so are at greatest risk, while others can be kept offline or on private servers and so are much harder to compromise.

Attacks

TUF provides protection against a number of attacks. If you want to imagine a concrete example, consider that the attacker runs or has compromised a mirror of Hackage that you use, and then think about each of these attacks:

Interestingly, a naive approach based only on authors signing packages only solves the first of the attacks above. All the others are still possible, and several of them can lead directly to compromised client systems, while others can force you to use old possibly vulnerable software, or just DOS your system.

Roles for keys

TUF defines a number of roles for keys:

Client verification process

In a full implementation, including delegated target keys, the client download and verification process goes like this:

At this point we know (with cryptographic evidence) whether the package index has changed and what size and hash to expect when we download it. We also have a bound on the freshness of this information because the timestamp signature includes a short (e.g. day or two) expiry time.

At this point we have downloaded the package index successfully and know (again with cryptographic evidence) that it has not been tampered with since being created on the server and so we can trust all the metadata in the index.

At this point we have downloaded an individual package and we know (with cryptographic evidence) that the package tarball has not been modified since it was created on the author’s system, and that this author is authorised to provide this named package.

Discussion

The different keys provide assurance of slightly different (but somewhat overlapping) things. If we only care about man-in-the-middle attacks then we only need the timestamp key and signature, since it includes the hash of the index which includes hashes of the packages.

The addition of the snapshot key to sign the index allows the snapshot key to be kept on a different system from the public-facing index server (think of an organisation like Red Hat with an internal system only available to people uploading packages). This timestamp/snapshot separation would allow the public-facing index server to be compromised with the damage being limited to freeze attacks (and the duration of this is bounded).

The addition of the target key (and delegated target keys) gives additional protection in the case that the index server is compromised. When target keys are being used, an attacker that has taken control of the index server can perform freeze attacks, rollback and mix and match attacks, but cannot change the contents of individual packages, nor provide packages that they are not authorised to provide.

Again, it is important to emphasise that all the keys need to be kept in separate places for them to have any security benefit.

Note that slow download and endless data attacks can be prevented because except for the initial tiny timestamp file, the client knows (securely) the expected download size and so carefully written download code can at least notice slow download and endless data attacks.

The Update Framework for Hackage

TUF is an architecture (albeit with a reference implementation) designed to be flexible enough to cover many use cases (distros, individual apps etc) and so we are free to adapt it for our particular use case.

For the case of Hackage as currently designed, the timestamp/snapshot key distinction does not add anything, as both keys would need to be on the same system. This is because the Hackage server is public facing both for download and upload. There would need to be a separate more secure upload server for the timestamp/snapshot distinction to add value. So for the Hackage implementation we will merge the timestamp and snapshot keys into a single package index signing key.

We also need to adapt the format and layout of the various bits of TUF metadata to fit the existing Hackage index format. For the most part this is fairly straightforward. The timestamp file can be provided as a separate file over HTTP, since it needs to be small and downloaded before the package index. All the other files that TUF specifies can be included inside the Hackage index file. This is straightforward because the Hackage index is simply a tarball and is designed to be extensible with extra file entries.

Our development plan had us start with the package index signing but keeping in mind the target keys feature so that it can be added later without any compatibility problems. We have enough funding to complete the package index signing and we hope that with this clear design roadmap we can secure either further funding or community volunteer effort.

Implementing target keys later would involve client tools (like cabal) generating keys and an out-of-band mechanism to submit them to be signed by the holder of the central target key (in practice the Hackage admins). For updating who is allowed to upload which package, a semi-manual process will be needed for the central target key holders to check and authorise re-signing the key delegation information. Tools like cabal would also need to be extended to sign tarballs and upload the signatures. So there would be some extra work for package authors but it could be kept fairly minimal. The main cost would be that changing the set of maintainers for a package on Hackage would become an asynchronous process involving a Hackage admin re-signing something.

As a final point, Hackage allows package authors and Hackage trustees to edit .cabal file metadata after release. So when we add target keys then these parts of the index should be protected with the target keys, not just the main package index key.

Comparing with other approaches

HTTPS

HTTPS is often mentioned as a way to improve Hackage security. As the TUF papers make clear, relying on the public PKI for security leaves a lot to be desired, given all the well-documented problems with trusting all the public certificate authorities in the world. In addition, relying on HTTPS does not allow for untrusted mirrors, whereas TUF’s index signing approach does allow that. Finally, on a practical note HTTPS is unfortunately complicated to support sanely on all platforms (OpenSSL is horrible, and distributing it for platforms that don’t come with it is painful). By contrast, TUF only relies on a crypto hash and signature verification algorithm. There are many choices for these including good choices that are small enough to embed existing audited C code directly into a Haskell library or tool. We plan to use ed25519 and in particular the C implementation extracted from NaCl.

GnuPG

Another approach that has been discussed and had some development effort is one based on author signing with GnuPG by Nikita Karetnikov. It was designed to fit with the existing Hackage index format with signatures being included in the package index tarball. The primary difficulty with a GnuPG-based system is in establishing the web-of-trust to a sufficient degree that signature verification can be done fully automatically without supervision from the user.

More GnuPG

Another recent GnuPG-based proposal comes from Chris Done. This proposal is under active discussion so in the spirit of constructive criticism it’s worth comparing it with TUF.

The first general point is that TUF has been designed by academic experts in the subject, based both on research and existing real-world systems. Our crypto-humility should cover not just crypto algorithms but extend to whole system designs. As I mentioned at the beginning of this post, though my initial thoughts on package index signing looked quite similar to the root+snapshot key part of the TUF system, there are many extra details that TUF covers that I would not have considered. With these in mind we can already spot underspecified areas in the proposal. For example one part of the spec says “Downloads the index from the web server or loads from cache if not modified”, but this leaves open the question of how we know securely if the index has been modified or not. TUF addresses this with the signed timestamp file. This kind of detail is not something I would have considered either, but TUF identifies the problem (that relying on HTTP cache control headers would be vulnerable to MITM) and has a solution. Generally in a system designed by crypto amateurs like us we would expect to have such holes. So we can do better by picking an existing expert design. TUF seems like a reasonable choice since it is designed to be somewhat flexible to fit different situations, and it’s being considered for both Python and Ruby’s package systems.

More specifically, the GnuPG-based proposal relies on a web of trust in which individual end users have to decide who they trust and configure their tools appropriately. A more recent version of the proposal includes signed mappings from packages to signers, which are basically statements of “I trust so and so to upload package foo”, but you still need to trust the author of that statement.

The system seems to be designed primarily for enterprise users who are using a carefully selected set of packages where a level of manual supervision to decide which authors to trust is justified. By contrast TUF’s target key delegation system is designed to be fully automatic for end users.

With TUF’s target key delegation, the chain of trust is very clear and simple: the root key signs the target key, which signs delegated keys and delegation information, and the delegated keys sign individual packages.

I also think it’s unfortunate that the signatures are not integrated as part of the Hackage index as in Nikita’s system (though it is in keeping with the general 3rd party design). On the positive side, the system has the flexibility to have 3rd parties make claims of trustworthiness, such as “I have manually audited this code” or “My CI system says this builds ok without obviously attempting to attack the system”. That extra complexity adds to the amount of manual supervision needed and seems to have been left out of the latest version of the proposal.

Finally because the proposal has nothing corresponding to the TUF snapshot key signature then all the corresponding attacks are possible, including rollback, freeze and mix-and-match attacks.

A central authority

A suggestion that has come up in recent discussion is that there should be a central authority for who is allowed to upload which package and that that information should be transparent. The criticism is that while Hackage is currently that central authority, it does not generate any verifiable evidence.

With TUF the target key delegation information is precisely that cryptographic evidence. The holders of the target key is that central authority and the signed delegation information is the cryptographic evidence of who is allowed to upload what, and that information is distributed to all clients.

Incremental updates

There has been a design for incremental updates of the Hackage index floating around for some years. With the TUF approach of putting more signatures into the index the pain of slow Hackage index downloads would become critical so our design incorporates that existing design for incremental updates to the index. This means most index updates will be 10s of kilobytes rather than multiple megabytes. This is actually even better than rsync (and likely better than git). The trick is that the tar format was originally designed to be append only (for tape drives) and so if the server simply updates the index in an append only way then the clients only need to download the tail (with appropriate checks and fallback to a full update). Effectively the index becomes an append only transaction log of all the package metadata changes. This is also fully backwards compatible.

Conclusion

We think implementing TUF for Hackage will be a major improvement to security. The initial implementation without target keys will be completely transparent to both end users and authors, and then the later addition of per-author target keys will further improve security by guarding against certain attacks even if the central server is compromised.

We accept that although we are programming experts, we are not crypto experts. We are hopeful that by using TUF we are reusing an existing expert crypto design and will not fall into the traps that typically bedevil our fellow crypto amateurs.