the issue is the split invalidation path in the redis cache backend:

1. linkedBlobStore.Delete calls blobAccessController.Clear during repository delete handling.
2. cachedBlobStatter.Clear forwards that invalidation into the cache layer.
3. repositoryScopedRedisBlobDescriptorService.Clear checks that the digest is a member of repo a, but then only calls upstream.Clear.
4. upstream.Clear deletes the shared digest descriptor and does not remove the digest from the repository membership set for repo a.
5. when repo b later stats or gets the same digest, the shared descriptor is recreated.
6. repositoryScopedRedisBlobDescriptorService.Stat for repo a accepts the stale membership and now trusts the repopulated shared descriptor, restoring access in the repository that already deleted its link.

this creates a revocation gap at the repository boundary. the blob is briefly inaccessible from repo a right after delete, which confirms the backend link was removed, and then becomes accessible again only because stale redis membership survived while a peer repository repopulated the shared descriptor.

attack scenario

1. an operator runs distribution with storage.cache.blobdescriptor: redis and storage.delete.enabled: true.
2. the same digest exists in both repo a and repo b.
3. the operator deletes the blob from repo a and expects repository-local access to be revoked.
4. repo a correctly returns blob unknown immediately after the delete.
5. an anonymous or unprivileged user requests the same digest from repo b, which still legitimately owns it and repopulates the shared descriptor.
6. a later request for the digest from repo a succeeds again because stale repo-a membership was never revoked from redis.

PoC

attachment: poc.zip

the attached PoC is a deterministic integration harness using miniredis and the pinned distribution source tree.

steps to reproduce

canonical:

unzip -q -o poc.zip -d poc
cd poc
make canonical

expected output:

[CALLSITE_HIT]: repositoryScopedRedisBlobDescriptorService.Clear->upstream.Clear->repositoryScopedRedisBlobDescriptorService.Stat
[PROOF_MARKER]: repo_a_access_restored=true repo_a_delete_miss=true repo_b_peer_warm=true
[IMPACT_MARKER]: repo_a_post_delete_read=true confidentiality_boundary_broken=true

control:

unzip -q -o poc.zip -d poc
cd poc
make control

expected control output:

[CALLSITE_HIT]: repositoryScopedRedisBlobDescriptorService.Clear->repositoryScopedRedisBlobDescriptorService.Stat
[NC_MARKER]: repo_a_access_restored=false repo_b_peer_warm=true

expected vs actual

• expected: after repo a deletes its blob link, later reads from repo a should keep returning blob unknown even if repo b still references the same digest and warms cache state.
• actual: repo a first returns blob unknown, then repo b repopulates the shared descriptor, and repo a serves the deleted digest again through stale repo-scoped redis membership.

impact

the confirmed impact is repository-local confidentiality failure after explicit delete. an operator can remove sensitive content from repo a, observe revocation working immediately after the delete, and still have the same content become readable from repo a again as soon as repo b refreshes the shared descriptor for that digest.

this is not a claim about global blob deletion. the bounded claim is that repository-local revocation fails, which breaks the expectation that deleting a blob link from one repository prevents further reads from that repository.

remediation

the safest fix is to make redis invalidation revoke repo-scoped state together with the backend link deletion. in practice that means removing the digest from the repository membership set, deleting the repo-scoped descriptor hash, and keeping that cleanup atomic enough that peer-repository warming cannot restore access in the repository that already deleted its link.

poc.zip PR_DESCRIPTION.md attack_scenario.md