Published November 13, 2026
| Version 1.0
Report
Restricted
End-to-End Private AI Conformance Audit — Clawd TEE Gateway
Description
This report audits the Clawd TEE gateway (clawd-tee-gateway, from the public tea cut of clawd-tee) at source level against an end-to-end private AI standard requiring: (1) a GPU TEE model runtime; (2) model host proof binding the live workload to a TLS SPKI or E2EE public key the gateway can enforce; and (3) gateway proof — a published attestation report, keyset, signed receipts, and attested upstream sessions — under a hard rule that forwarding fails closed when no enforceable channel binding exists. Findings: the GPU TEE runtime is attested-not-bound (inference runs in the upstream's NVIDIA Confidential Compute, but the gateway checks nothing about it, and its own GPU evidence collection is a stub); model host proof is absent (the keyset's TLS and E2EE key lists are hardcoded empty; no SPKI pinning or upstream attestation fetch exists); gateway proof is mixed (the report, keyset, and signed receipts are genuinely published, with on-chain signature verification at level 1, but the quote is a self-labeled stub off TDX hardware and attested upstream sessions are not implemented). The headline result: forwarding is fail-open with honest labeling — verification failures degrade to log lines behind an HTTP 200, while receipts never falsely claim verification. The report scopes the path to enforced conformance per gap and proposes exact amendments.
Files
Additional details
Related works
- Is documented by
- Other: https://clawd-tee-gateway-4m6d.onrender.com (URL)
- Is supplement to
- Software: https://github.com/Solizardking/tea (URL)