Anthropic's FlashProve: Cryptographic AI Model Identity
Original: Anthropic is working on a system to prove AI models’ identity
Why This Matters
AI output authenticity is a growing trust problem across enterprise and regulatory contexts.
Anthropic published an open source system called FlashProve on Nov. 3, 2026, using cryptographic proofs to verify which specific AI model produced a given output, without exposing model weights.
Anthropic has released a technical specification and working code for FlashProve, an open source system designed to provide cryptographic guarantees—called 'provable inference'—that a specific AI model produced a given output. The concept is analogous to cryptographic email signatures: users can confirm an answer genuinely came from the model they think they used, even if the lab operating it tried to swap in a different one.
The researchers highlight two concrete risks the system addresses: AI companies secretly substituting cheaper models to cut costs, or deploying stronger models to inflate benchmark results. Third-party verification is another use case—a developer claiming their code was audited by a frontier cybersecurity model could now back that up cryptographically.
FlashProve works by requiring labs to publish a per-model 'fingerprint' that can be matched to outputs later. Critically, these fingerprints do not expose model weights—the proprietary parameters closed-model makers like Anthropic and OpenAI guard closely—making adoption by commercial labs plausible. The system is also designed to run locally on-device.
The work is explicitly framed as intermediate research, not a deployment-ready product. 'We are sharing this with the goal that others look at what we have done, take it, and improve it,' the team wrote. No timeline for deployment at Anthropic or any other lab has been announced.