Anthropic announces FlashProve system to verify AI model identity

Anthropic has announced FlashProve, an open source system that uses cryptographic proofs to verify which AI model generated a specific output. The system allows users to mathematically confirm that responses came from the claimed model, preventing scenarios where providers might secretly swap cheaper or more expensive models. While still in research phase with no deployment plans, the technical specification and code are publicly available on GitHub.

Key points

• FlashProve provides "provable inference" — a cryptographic guarantee that output came from a specific model, similar to email signatures that verify sender identity
• The system requires AI labs to publish a "fingerprint" of each model that can be matched to outputs without revealing proprietary model weights
• Designed to prevent model swapping, whether substituting cheaper models to cut costs or expensive ones to inflate performance results
• Can verify third-party claims, such as whether code was scanned by a frontier cybersecurity model or whether a business plan was reviewed by a specific model
• Verification is lightweight enough to run on local devices

How to use it

• AI labs can publish model fingerprints that enable verification for commercial models without exposing the weights that define model behavior
• Developers can verify that third-party services actually used the claimed model for tasks like vulnerability scanning or code review
• Users can confirm responses came from the model they're paying for rather than a cheaper substitute
• The open source code and detailed technical specification are available on GitHub for researchers and developers to examine and improve upon

Could model authentication become a standard feature users expect from AI providers?