Anthropic 的 FlashProve 验证 AI 模型响应，无需暴露权重
anthropic 发布 flashprove，一种使用加密技术验证 ai 响应的方法，以解决模型替换和数据篡改问题。这影响了 AI 模型提供商和监管机构。
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.