Anthropic open-sources FlashProve to cryptographically verify AI model identity
FlashProve provides 'provable inference' to confirm outputs match claimed models without exposing proprietary weights, though deployment plans remain unspecified.
1 source · cross-referenced
- Anthropic released FlashProve, an open-source system for cryptographically verifying which AI model produced a given output.
- The system uses model 'fingerprints' to confirm identity without revealing internal weights, enabling verification for closed models.
- FlashProve is designed to prevent model substitution by labs or third parties, with lightweight verification possible on local devices.
- Code and technical specifications are available on GitHub, but Anthropic has not announced plans to deploy the system internally or externally.
- Researchers emphasize this is intermediate progress and invite public improvements to the approach.
Anthropic published a blog post describing FlashProve, an open-source system intended to provide "provable inference" for AI model outputs. The system generates cryptographic proofs that a specific output originated from a claimed model, analogous to email signatures that verify a sender’s identity.
The system requires AI labs to publish a unique "fingerprint" for each model release, which can later be matched to outputs and verified without exposing the model’s proprietary weights. This design allows closed models to participate without disclosing internal architecture or parameters.
FlashProve is positioned to address scenarios where a model provider might substitute a cheaper or more expensive model than advertised, or where third parties claim unverifiable use of advanced models for tasks like cybersecurity scanning or business planning.
Verification is designed to be lightweight enough to run on local devices, supporting decentralized and user-controlled validation of model outputs.
Anthropic released detailed technical specifications and working code under an open-source license on GitHub, framing the work as intermediate research rather than a completed agenda.
The company stated there are no current plans to deploy FlashProve at Anthropic or any other lab, and invited external researchers to review, use, and improve the system.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79