Anthropic Integrates Crypto Sector’s Zero-Knowledge Proofs into AI Inference to Prevent Secret Model Swapping.

Beating AI News Flash: Anthropic has open-sourced FlashProve, bringing zkSNARK (zero-knowledge proof) technology—commonly used in the cryptocurrency sector—into AI model inference. Originally, this technology was used to validate blockchain transactions without disclosing transaction details. Anthropic aims to leverage it to solve another problem: when users pay premium prices for an AI model’s API, how can they confirm the service provider has not secretly swapped it for a cheaper alternative? FlashProve enables providers to first generate an encrypted fingerprint for model weights, then issue a mathematical proof for inference outputs. Users do not need to access the model weights or re-run the model to verify if a given output comes from the pre-committed set of weights. Claude also participated in the system’s research, development and optimization. Anthropic has completed experiments on Meta’s 400-billion-parameter model Llama 4 Maverick. For every 1024-token context processed, the generated proof is approximately 19 MB, and verification on a standard CPU takes around 17 seconds. However, each proof currently can only verify one output token, and cannot validate an entire response. The GPU compute time required to generate the proof is roughly 70 to 95 times that of standard inference. Whether it can eventually be used to root out API providers that secretly swap models will depend on when proof costs can be reduced.