Relay Attack Vulnerability in Cocos AI Confidential Computing System by Ultravioletrs
CVE-2026-33697
What is CVE-2026-33697?
The current implementation of attested TLS in Cocos AI is vulnerable to a relay attack, affecting versions from v0.4.0 to v0.8.2. This vulnerability allows an attacker to extract the ephemeral TLS private key used during intra-handshake attestation, which can be exploited to impersonate an attested Cocos AI service. As a result, clients may unknowingly accept connections under false assumptions, undermining the authentication guarantees intended by attested TLS. The attacker may exploit this weakness via physical access to hardware or using transient execution and side-channel attacks. Although the aTLS implementation underwent significant redesign in v0.7.0, the inherent architectural flaws mean that this vulnerability is still present. Currently, no patches or complete workarounds exist, but implementing hardening measures such as updating TEE firmware, enforcing strict attestation policies, and enabling mutual aTLS with CA-signed certificates can help mitigate the risk.
Affected Version(s)
cocos >= 0.4.0, < 0.8.2
News Articles
Confidential computing's core trust mechanism is broken. The fix may not exist
Attested TLS: the handshake that can't prove who's on the other end