Anthropic's FlashProve aims to cryptographically verify which AI model produced an output

FlashProve arrives as a response to a practical trust problem: users of AI services generally have no way to confirm that the model answering them is the one advertised. Anthropic's researchers frame the system as analogous to cryptographic email signatures, and they describe scenarios where a developer claims a software library was scanned by a frontier cybersecurity model, or a company claims its business plan was approved by an appropriate model, asking how a user could verify such claims.

The design choice to keep model weights private is central. Because the published fingerprints do not expose weights, labs could in principle participate without surrendering their most valuable asset, which is what would allow the system to extend to commercial models. The researchers note that FlashProve is meant to be robust even against the lab that built and runs the model, addressing cases where an AI company might quietly substitute a cheaper model to cut costs or a more expensive one to inflate performance results.

Anthropic is explicit that this is intermediate work. The announcement includes a detailed technical specification and working code for various elements on GitHub, but the researchers say there are no plans to deploy it at Anthropic or any other lab, and they describe the post as a way to share progress so others can take it and improve it.