Anthropic's research team has opened-source FlashProve, using cryptographic proofs to verify the identity of AI model outputs.

Anthropic has released a research blog post introducing FlashProve, an open-source system that uses cryptographic proofs to verify whether a specific output came from a particular model. The technical specifications and code are now publicly available on GitHub. The system requires model publishers to disclose non-weight-leaking model fingerprints, allowing verification to be performed locally on devices. This can help prevent labs from secretly swapping in cheaper or more expensive models, though there are currently no plans to deploy it at Anthropic or other labs.