Auditing Anonymous AI Models: A Four-Stage Protocol for Black-Box Identity Verification
The paper offers a black-box audit protocol for identifying anonymous API-served AI models without trusting their self-claims.
Yisen Xi describes four stages: reconstructing launch configuration from archived snapshots, matching configuration against catalogs, testing tokenizer identity, and checking behavior with probes. The validation is limited: the paper reports consistency tests on 10 known releases, not full anonymous identification at scale. It also cites one prospective flagship case where the protocol inferred the GLM-5.3 family/version line before an official reveal. ArXiv · AI/CL/LG's note
Yisen Xi describes four stages: reconstructing launch configuration from archived snapshots, matching configuration against catalogs, testing tokenizer identity, and checking behavior with probes. The validation is limited: the paper reports consistency tests on 10 known releases, not full anonymous identification at scale. It also cites one prospective flagship case where the protocol inferred the GLM-5.3 family/version line before an official reveal. ArXiv · AI/CL/LG's note
score 4