Log verification for expired targets

[haydentherapper]

Overview

This is a tracking issue for supporting verification for expired/rotated targets. @asraa and I will be working on this.

Currently, cosign assumes the latest TUF metadata can be used to validate signatures. As the Fulcio CA certificate will expire, we will have to rotate that target at some point. This will cause cosign to not be able to validate the signatures that chain up to the expired CA certificate, since the TUF metadata will contain a different certificate.

We propose bundling a pointer to the metadata used when generating the signature. We will do this by including the snapshot or timestamp JSON in the signature bundle. Cosign will use this to find versioned TUF metadata.

Design doc

Tasks

For cosign:

• Add current snapshot/timestamp.json to an entry’s bundle
• Add verification by pulling the referenced versioned x.targets.json and x.root.json
• Cache previous x.targets.json files and expired Fulcio certificates

There are a few tasks for the Sigstore TUF repo.

• Enable consistent snapshots in Sigstore root
• Add versioned targets.json and complete list of targets to Sigstore’s TUF root (and synced to GCS bucket)
• Generate snapshot/timestamp more frequently

[asraa]

Some more thoughts to make the experience of expired target verification along with BYO TUF root:

• we may want to change the sigstore TUF root to pull in directories of targets rather than hard-coded target names (e.g. add all targets in a path fulcio/** to the trusted root pool rather than pull in a hardcoded target name fulcio_v1.crt.pem). this would reduce friction between any target version names and BYO TUF people

[haydentherapper]

A few other things we'll need to consider:

• I need to update cosign attest to store the timestamp too, this should be a simple change.
• How does this work with offline verification when you provide a locally saved bundle? We'll have to look in the cache of saved targets and fail if the target that's referenced in the timestamp is missing.
• cosign sign-blob is a little trickier, since it only outputs a signature. Priya is working on saving a local bundle, which will include the timestamp which we can use. For online verification, maybe we provide the timestamp with the signature? Or maybe we deprecate returning only the signature and return the bundle?

[haydentherapper]

As I was implementing verification using the TUF timestamp, I ran into an issue. As a reminder, the verification flow is roughly:

1. Read the timestamp from an annotation on the OCI image
2. Using the timestamp, fetch the versioned snapshot
3. Using the snapshot, fetch the versioned root (N) and targets
4. Verify the root by walking the chain from 1.root.json to N.root.json
5. Verify the timestamp, snapshot and targets
6. Use the old targets to verify the OCI image

The TUF spec no longer requires that the root metadata be included with the snapshot. Recently, go-tuf removed including root.json in the snapshot. This means step (3) is not possible, because there will be no reference to a versioned root.

This leaves us a few options:

• Keep generating snapshots that contain the root, ignoring the spec. I am concerned this option will cause issues in the future.
• Attempt verification using all root metadata that could be valid when the signature was generated. This sounds like an overly complex solution. A bit too wide of a cast net.
• Implement another solution we had proposed, storing all targets. See the design doc for more details.

To summarize, this solution is far simpler. On verification, we will simply load all targets from the target metadata. For the verification of the signature, we will load all Fulcio certs into the root pool. We will have something similar for bundle verification using the Rekor public keys. We'll keep track in custom TUF metadata of which target is active, and use that to inform users when they're verifying a signature using old targets.

The main trade-off is that the targets metadata will grow over time. If we want to cap its size, we will need to decide on how long we'll keep old targets around. This can be decided later though, as we aren't frequently rotating targets.

This also simplifies the revocation story. To mark a target as revoked, we simply remove it from the set of targets, which is aligned with how TUF expects revocation to occur. @asraa and I do think there's still value in tracking revocation somehow, through a TUF delegation - I have a separate doc discussing this in more detail that I'll circulate shortly.

@dlorenc @bobcallaway - Y'all reviewed the original proposal, lemme know if you have any concerns with this other approach.