Anthropic brings cryptocurrency zero-knowledge proofs into AI inference to prevent models from being secretly 'swapped'

Beating AI News Flash: Anthropic has open-sourced FlashProve, bringing zkSNARK (zero-knowledge proof) technology commonly used in the cryptocurrency space to AI model inference.

This technology was originally designed to prove the validity of blockchain transactions without revealing transaction details. Anthropic wants to use it to solve a different problem: when users pay a premium for an AI model's API, how do they know whether the service provider has secretly swapped in a cheaper model?

FlashProve has service providers first generate cryptographic fingerprints for model weights, then provide mathematical proofs for inference results. Users can verify whether a given output came from the pre-committed set of weights without obtaining the model weights or re-running the model. Claude also participated 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 1,024 tokens of context processed, the generated proof is approximately 19 MB, and verification on a standard CPU takes about 17 seconds.

However, each proof can currently only verify a single output token and cannot yet validate an entire response. The GPU compute time required to generate proofs also reaches approximately 70–95 times that of normal inference. Whether this can eventually catch API service providers secretly swapping models will depend on when proof costs come down.