Anthropic launches open-source FlashProve for AI model verification

Anthropic has released FlashProve, an open-source system designed to cryptographically verify which AI model produced a given output. The AI company detailed the system in a blog post published on Tuesday, open-sourcing its technical specifications and code on GitHub.

FlashProve provides a framework for AI model verification and output authenticity. The system's core goal is what the company calls 'provable inference.' This creates a cryptographic proof that mathematically guarantees an output originated from a specific model.

Cryptographic Proof for Model Identity

FlashProve enables provable inference through cryptographic proofs. These proofs offer a mathematical guarantee that an output came from a specific model. The system allows for the verification of third-party claims about the execution of complex tasks. This could include claims that an advanced model performed specific work like vulnerability scanning.

Preventing Model Substitution and Fraud

The system directly addresses risks in commercial AI deployment. FlashProve could prevent AI companies from secretly substituting models. Cheaper models might be used to cut costs, while more expensive ones could be swapped in to inflate performance results. The technology enables verifiable inference to prevent such model substitution in commercial operations. It provides a check against fraud.

Privacy-Preserving Verification

FlashProve's method protects proprietary information. It requires AI labs to publish a cryptographic fingerprint for each model. This fingerprint can be matched to outputs and verified without revealing the model's internal weights. The approach allows labs to share fingerprints for verification while keeping their core model assets private. This is key for verifying commercial models without compromising trade secrets.

Local Device Compatibility

Verification is designed to be lightweight and decentralized. FlashProve is built to be lightweight enough for verification to occur directly on a local device. This support for local verification means the process does not necessarily require cloud infrastructure.

Researchers released a detailed technical specification and working code for FlashProve under an open source license. The research is ongoing, with no current plans to deploy the system at Anthropic or any other lab. The system is engineered to be robust against impostors, including the model's own creator, attempting to disguise one model as another.

The announcement clarifies that the blog post shares intermediate progress, not a completed agenda. The team encourages others in the community to review and improve upon the open-source code. FlashProve is intended for sharing milestone results and providing an industry reference, with no specific commercial deployment plans outlined.