Anthropic's Frontier Red Team released FlashProve providing cryptographic proofs that an AI response originated from committed model weights