A new method lets outsiders test whether an AI service is really running the model it claims to be.
Researchers describe the approach, called TP-CRIV (Third-Party Challenge-Response Identity Verification), in a paper posted to arXiv (arXiv:2609.29264) on September 25, 2026. The framework lets a verifier probe a deployed AI service through its normal black-box API, without special access or cooperation from the provider, and issue fresh challenge questions the system could not have prepared for in advance. Network isolation during testing rules out the service quietly fetching help from elsewhere after the challenge is issued. The team tested the method on ten ImageNet-pretrained TorchVision image classifiers and found it reliably distinguished matching models from non-matching ones using calibrated statistical thresholds.
Model misappropriation, deploying someone else's AI model and passing it off as your own, is getting harder to catch as more products run behind opaque APIs. Existing tools like watermarking and fingerprinting mostly compare outputs or rely on evidence baked in ahead of time; TP-CRIV instead asks a live pass/fail question of whatever is running right now. That matters for licensing disputes and audits where nobody involved gets to peek under the hood.
This is an early result on a narrow slice of models, image classifiers, not the large language models most of the AI industry actually argues about, so read it as a promising method rather than a ready-made policing tool.