Anthropic open-sources FlashProve to check which AI model made an output

Anthropic is building an open-source system called FlashProve, aiming for 'provable inference': using cryptographic proofs to mathematically confirm an AI output really came from a specific model. It works like encrypted email signatures, blocking model impersonation—even by the company running the model itself.

The system can stop AI companies from quietly swapping in cheaper models to cut costs, or pricier ones to inflate benchmark scores. It also helps verify third-party claims that a task was done by an advanced model. Labs give each model a 'fingerprint' that hides the model weights, so closed-source models can join in too. It's lightweight, checks run on local devices, and the code is open on GitHub.

The research is still ongoing, and Anthropic has no plans to deploy it yet. The release invites others to review, build on, and improve the work.