Anthropic and academic researchers have launched FlashProve, an open-source cryptographic tool using zero-knowledge proofs. The system allows AI providers to prove that a specific language model generated a response without revealing the model's proprietary weights.