Anthropic's FlashProve verifies AI model responses without exposing weights
anthropic releases flashprove, a research prototype that uses cryptography to verify ai responses, tackling the problem of model swapping and data tampering. this affects ai model providers and regulators.
anthropic released flashprove, a research prototype that verifies ai model responses using cryptography. it creates a digital fingerprint of the model's weights and delivers a compact proof with every response, revealing nothing about the secret model weights. flashprove generates proofs at 110 tokens per second on a single H100 GPU for meta's llama 4 maverick model. regulators and companies can use this system to verify model usage and detect tampering. however, flashprove has limitations, including context window and sampling limits.