Anthropic has introduced FlashProve, a research system designed to verify the origin of AI model outputs using cryptographic proofs. The mechanism allows users to confirm that a response was generated by a specific model rather than a substitute, preventing providers from silently switching to cheaper or different alternatives.
The system functions similarly to digital signatures in email. Developers publish cryptographic fingerprints of their models, which can be compared against outputs to verify authenticity. These fingerprints do not expose the model's weights, making the tool suitable for proprietary systems with closed architectures.
FlashProve enables verification of claims regarding which model performed a specific task, such as code security audits or business plan evaluations. The verification process is designed to be lightweight, allowing it to run on local devices without significant performance overhead.
Anthropic has released the technical specifications and source code for FlashProve components on GitHub under an open license. The company currently has no plans to implement the system in its own products or other laboratories, as it is intended for community review and improvement.
