Anthropic Introduces FlashProve: Cryptographic Verification System for Model Inference Authenticity
Anthropic's Frontier Red Team released FlashProve, an open-source zkSNARK system enabling cryptographic proofs that API responses originate from a specific committed model without revealing weights. The system achieves approximately 110 tokens per second on a single H100 GPU when
Anthropic's Frontier Red Team released FlashProve, an open-source zkSNARK system enabling cryptographic proofs that API responses originate from a specific committed model without revealing weights. The system achieves approximately 110 tokens per second on a single H100 GPU when processing a 400B mixture-of-experts model, with 18.78 MB proofs verifiable on CPU in 17 seconds. FlashProve addresses the inability of users to verify which model generated their API responses, supporting use cases including regulatory compliance, tamper detection, and text provenance verification. The protocol is formally verified in Lean for security properties including completeness and soundness. Current limitations include a 1,024 token context window and support restricted to mixture-of-experts architectures.