Anthropic FlashProve targets secret AI model swapping
The open-source FlashProve project would let users check that an AI answer really came from the model a company claims to be running, without exposing private model weights.
Anthropic researchers are building FlashProve, an open-source system for “provable inference” of which AI model produced a response.
Labs would publish a cryptographic fingerprint for each model so outputs can be checked without revealing private weights.
The design aims to deter secret swaps of cheaper or stronger models and to let third parties prove a named model did a task.
A technical specification and working code are on GitHub; researchers say there are no deployment plans at Anthropic or other labs yet.