Anthropic applies zero-knowledge proof technology to AI reasoning to prevent model swapping

Anthropic has open-sourced FlashProve, bringing the commonly used zkSNARK (zero-knowledge proof) technology from the cryptocurrency field into AI model inference.

This technology was originally designed to prove the validity of blockchain transactions without disclosing transaction details. Anthropic aims to use it to solve another problem: when users pay a high price for an AI model's API, how can they know if the service provider has secretly switched to a cheaper model?

FlashProve allows service providers to first generate encrypted fingerprints for the model weights and then provide mathematical proof for the inference results. Users do not need to obtain the model weights or rerun the model; they can check whether the specified output comes from the previously committed set of weights. Claude was also involved in the development and optimization of this system.

Anthropic has already completed experiments on Meta's 400 billion parameter model Llama 4 Maverick. For every 1024 tokens of context processed, the generated proof is about 19 MB, and verification on a regular CPU takes about 17 seconds.

However, currently, each proof can only verify one output token and cannot validate an entire response. The GPU computation time required to generate the proof is also about 70 to 95 times that of regular inference. Whether it can be used to identify API service providers that secretly switch models in the future will depend on when the cost of proof generation can be reduced.