Anthropic develops new open-source system to verify AI model identities

AI startup Anthropic is building an open-source system called FlashProve to prove which specific artificial intelligence model generated a particular output. Researchers revealed details of the project in a blog post published on Tuesday, explaining that the tool provides what they call provable inference through mathematical guarantees.

How It Works

The system works similarly to cryptographic email signatures by letting users confirm that an answer truly came from the claimed sender. FlashProve stops impostor labs from disguising one model as another to cut costs with cheaper options or inflate performance with expensive ones. AI labs would publish a public fingerprint for each model that matches outputs without exposing private internal weights. This lightweight design allows users to run verifications directly on a local device.

Current Status

Researchers have released technical specifications and working code on GitHub under an open-source license. The project remains an ongoing research effort with no immediate plans for deployment at Anthropic or other labs. The team shared this progress to encourage outside developers to review and improve the technology.

Context

Cryptographic verification methods are widely used in digital security to prove authenticity without revealing underlying secret data. As AI models become more common in commercial software, verifying their true origins helps prevent misleading performance claims.