• Sources: primary, discussion
  • Summary: Anthropic's Frontier Red Team published FlashProve on 2026-11-02. It is a zero-knowledge proof system that binds each response to a committed set of weights and hides those weights but not the model architecture. Proving costs about 72 to 94 times the GPU time of vLLM. The release includes the prover, a C++ verifier, a Lean verifier and Lean proofs of the top-level security statements.
  • Why it matters: It offers a software-only way to check that a router or provider served the committed model, though the 70x to 95x overhead limits it to audits for now.
  • Comments: Hacker News readers focus on the cost overhead and on whether third-party routers would ever adopt such proofs.