AI
Anthropic develops FlashProve to cryptographically prove AI output came from a specific model
Anthropic has released an open-source research project called FlashProve, aimed at verifying through cryptographic proofs that a piece of AI output was generated by a specific model. The system requires model labs to release model “fingerprints” that do not reveal the weights, and it can be verified locally with little overhead. It remains in the research stage, with no deployment plans for now.
According to TechCrunch, Anthropic introduced an open-source system called FlashProve in a blog post on Tuesday. The system is designed to enable what researchers call “provable inference”: using cryptographic proofs to provide a mathematical guarantee for a piece of output, confirming that it indeed came from the specific model the user believes it came from.
FlashProve works similarly to an encrypted email signature and can resist attempts to disguise one model as another, even if the party carrying out the disguise is the lab that built and operates the model. Researchers said, for example, that it could prevent AI companies from secretly using a cheaper model to cut costs, or using a more expensive model to exaggerate performance. It could also verify third-party claims that tasks such as vulnerability scanning and business-plan assessment were performed by an advanced model.
The system requires AI labs to publish a “fingerprint” for every released model that can be matched with subsequent output. Crucially, these fingerprints do not expose model weights, so closed-source commercial models can also participate. The verification is designed to be lightweight and can be completed on a local device. Anthropic also released technical specifications and partial working code, but the research is still ongoing, and neither the company nor other labs have deployment arrangements for now.