Putting Digital Fingerprints on AI Outputs: Anthropic Develops FlashProve to Leave Model Swapping Nowhere to Hide
Anthropic is developing an open-source verification system called FlashProve, aimed at addressing whether the AI answers users receive are truly generated by the model they claim to be. In a Tuesday blog post, the company said its 'provable inference' will use cryptography to provide a mathematical guarantee for every output, confirming that it came from a specified model, similar to how cryptographic email signatures verify the sender. The system can prevent model A from masquerading as model B, close a hidden backdoor for cost cheating, and expose inflated benchmark results, striking directly at a long-standing unspoken practice in the industry.