Source · Tier C · Blog / article

Auditing a Frontier Model Without Seeing its Weights

D. McCann-Sayles, T. Verma. 2026. Tinfoil blog.

Originalhttps://tinfoil.sh/blog/2026-06-09-auditing-frontier-model-without-weights
Accessed2026-09-25
Authored or published byTinfoil
NoteTinfoil's account of Pour Demain running interpretability evaluations (vllm-lens) on GLM-5.1, about 744 billion parameters, on eight H200 GPUs in Tinfoil's confidential-computing platform, with the lab supplying the model, the auditor its evaluation code, and the weights kept inside the enclave. Reports 33-38% overhead with interpretability hooks enabled. Supports only "Tinfoil reports" statements.

Cited by

Search

Full search page