Matthew Garrett

Last month I wrote about how bearer tokens are just awful, and a week later Github announced that someone had managed to exfiltrate bearer tokens from Heroku that gave them access to, well, a lot of Github repositories. This has inevitably resulted in a whole bunch of discussion about a number of things, but people seem to be largely ignoring the fundamental issue that maybe we just shouldn't have magical blobs that grant you access to basically everything even if you've copied them from a legitimate holder to Honest John's Totally Legitimate API Consumer.

To make it clearer what the problem is here, let's use an analogy. You have a safety deposit box. To gain access to it, you simply need to be able to open it with a key you were given. Anyone who turns up with the key can open the box and do whatever they want with the contents. Unfortunately, the key is extremely easy to copy - anyone who is able to get hold of your keyring for a moment is in a position to duplicate it, and then they have access to the box. Wouldn't it be better if something could be done to ensure that whoever showed up with a working key was someone who was actually authorised to have that key?

To achieve that we need some way to verify the identity of the person holding the key. In the physical world we have a range of ways to achieve this, from simply checking whether someone has a piece of ID that associates them with the safety deposit box all the way up to invasive biometric measurements that supposedly verify that they're definitely the same person. But computers don't have passports or fingerprints, so we need another way to identify them.

When you open a browser and try to connect to your bank, the bank's website provides a TLS certificate that lets your browser know that you're talking to your bank instead of someone pretending to be your bank. The spec allows this to be a bi-directional transaction - you can also prove your identity to the remote website. This is referred to as "mutual TLS", or mTLS, and a successful mTLS transaction ends up with both ends knowing who they're talking to, as long as they have a reason to trust the certificate they were presented with.

That's actually a pretty big constraint! We have a reasonable model for the server - it's something that's issued by a trusted third party and it's tied to the DNS name for the server in question. Clients don't tend to have stable DNS identity, and that makes the entire thing sort of awkward. But, thankfully, maybe we don't need to? We don't need the client to be able to prove its identity to arbitrary third party sites here - we just need the client to be able to prove it's a legitimate holder of whichever bearer token it's presenting to that site. And that's a much easier problem.

Here's the simple solution - clients generate a TLS cert. This can be self-signed, because all we want to do here is be able to verify whether the machine talking to us is the same one that had a token issued to it. The client contacts a service that's going to give it a bearer token. The service requests mTLS auth without being picky about the certificate that's presented. The service embeds a hash of that certificate in the token before handing it back to the client. Whenever the client presents that token to any other service, the service ensures that the mTLS cert the client presented matches the hash in the bearer token. Copy the token without copying the mTLS certificate and the token gets rejected. Hurrah hurrah hats for everyone.

Well except for the obvious problem that if you're in a position to exfiltrate the bearer tokens you can probably just steal the client certificates and keys as well, and now you can pretend to be the original client and this is not adding much additional security. Fortunately pretty much everything we care about has the ability to store the private half of an asymmetric key in hardware (TPMs on Linux and Windows systems, the Secure Enclave on Macs and iPhones, either a piece of magical hardware or Trustzone on Android) in a way that avoids anyone being able to just steal the key.

How do we know that the key is actually in hardware? Here's the fun bit - it doesn't matter. If you're issuing a bearer token to a system then you're already asserting that the system is trusted. If the system is lying to you about whether or not the key it's presenting is hardware-backed then you've already lost. If it lied and the system is later compromised then sure all your apes get stolen, but maybe don't run systems that lie and avoid that situation as a result?

Anyway. This is covered in RFC 8705 so why aren't we all doing this already? From the client side, the largest generic issue is that TPMs are astonishingly slow in comparison to doing a TLS handshake on the CPU. RSA signing operations on TPMs can take around half a second, which doesn't sound too bad, except your browser is probably establishing multiple TLS connections to subdomains on the site it's connecting to and performance is going to tank. Fixing this involves doing whatever's necessary to convince the browser to pipe everything over a single TLS connection, and that's just not really where the web is right at the moment. Using EC keys instead helps a lot (~0.1 seconds per signature on modern TPMs), but it's still going to be a bottleneck.

The other problem, of course, is that ecosystem support for hardware-backed certificates is just awful. Windows lets you stick them into the standard platform certificate store, but the docs for this are hidden in a random PDF in a Github repo. Macs require you to do some weird bridging between the Secure Enclave API and the keychain API. Linux? Well, the standard answer is to do PKCS#11, and I have literally never met anybody who likes PKCS#11 and I have spent a bunch of time in standards meetings with the sort of people you might expect to like PKCS#11 and even they don't like it. It turns out that loading a bunch of random C bullshit that has strong feelings about function pointers into your security critical process is not necessarily something that is going to improve your quality of life, so instead you should use something like this and just have enough C to bridge to a language that isn't secretly plotting to kill your pets the moment you turn your back.

And, uh, obviously none of this matters at all unless people actually support it. Github has no support at all for validating the identity of whoever holds a bearer token. Most issuers of bearer tokens have no support for embedding holder identity into the token. This is not good! As of last week, all three of the big cloud providers support virtualised TPMs in their VMs - we should be running CI on systems that can do that, and tying any issued tokens to the VMs that are supposed to be making use of them.

So sure this isn't trivial. But it's also not impossible, and making this stuff work would improve the security of, well, everything. We literally have the technology to prevent attacks like Github suffered. What do we have to do to get people to actually start working on implementing that?