GPU VulnDB

Database/AI/ML frameworks & serving

Cocos AI: TDX aTLS verifier ignores expected REPORT_DATA, accepting replayed attestation evidence

CVSS 9.1CVE-2026-92701AI/ML frameworks & servingcurated

Impact

Cocos AI is used to run AI workloads inside confidential VMs, so the attestation check is the only thing proving that the enclave a relying party is about to hand a model or dataset to is the one it thinks it is. In 0.8.2 and earlier the TDX verifier receives the expected current-session value but never applies it to TdQuoteBodyPolicy.ReportData, so a structurally valid TDX QuoteV4 with mismatched or reused REPORT_DATA passes. An attacker who can present previously captured evidence - or evidence from an unrelated TEE instance - can get the relying party to release application data or grant authorization for an attestation context that is not the current TLS session. On a shared confidential-AI fleet this collapses the tenant separation the TEE was bought for: model weights and input data can be released to a session that was never actually attested.

Who can reach it

Network, no authentication: anyone who can complete the aTLS handshake with a relying party that uses this verifier, holding any structurally valid TDX quote (including a replayed one).

What to do

Upgrade Cocos AI to 0.9.0, which removes the affected verifier path as part of the move to post-handshake attestation. This is a control-plane/agent upgrade plus a service restart on the relying-party side; the TDX guests themselves do not need reprovisioning, but any trust decision made by an 0.8.2-or-earlier relying party should be treated as unverified. The advisory names no backport to the 0.8.x line.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.