Auditing Anonymous AI Models: A Four-Stage Protocol for Black-Box Identity Verification
A 2026 arXiv paper proposes a four-stage forensic audit protocol for black-box identity verification of API-served anonymous AI models. Stage 0 reconstructs launch-time configuration from archived platform snapshots; Stage 1 fingerprints configuration against the platform catalog; Stage 2 tests tokenizer identity with a cross-length differential; Stage 3 corroborates with behavioral probes. The protocol was tested on 10 known-identity releases, achieving 7 exact matches, 2 precision differences, 1 partial, and 0 counter-directional results. A prospective validation on a flagship case pointed to the GLM-5.3 version.
The paper addresses the rise of stealth AI model releases on developer platforms and the lack of validated methods for black-box identity verification. It introduces a four-stage protocol combining archived configuration reconstruction, configuration fingerprinting, tokenizer differential testing, and behavioral probes. Testing on 10 known-identity releases showed mostly exact or precision-difference matches, with no counter-directional results. A prospective case study identified a model as GLM-5.3.
The protocol's Stage 2 tokenizer identity test uses a cross-length differential to reject short-prompt collisions, improving robustness over naive tokenizer matching. Stage 0's use of Internet Archive snapshots enables detection of preview-production drift, a novel forensic angle for API-served models.
The work responds to a market need for supply-chain risk assessment and capability expectations when using anonymous models. It provides a systematic alternative to practitioner checklists, which lack accuracy evidence, and could become a standard for model auditing.
Enables enterprises to verify the identity of third-party AI models, reducing supply-chain risk and ensuring compliance with data-handling terms. Could support due diligence in model procurement and monitoring of model drift.
Next signals include adoption of the protocol by AI auditing firms, publication of end-to-end identification results under anonymity, and potential integration into model registry or compliance frameworks.