Anthropic introduces FlashProve to combat model counterfeiting using zero-knowledge proofs

Anthropic has partnered with zero-knowledge proof researchers to release FlashProve, a system designed to authenticate large language model weights and combat the gray market of model wrapping. The verification process involves generating a public fingerprint ID for the weights, which takes approximately 39 minutes on a single H100 GPU before deployment. Following each interaction, the GPU produces a proof of about 19 MB, which an ordinary CPU can validate in 17 seconds without needing the model weights. To handle massive models like Llama 4 Maverick (400 billion parameters, 232 GB in 4-bit weights), FlashProve utilizes Mixture-of-Experts (MoE) architecture. This allows the proof size to relate only to the 17 billion activated parameters per token, effectively bypassing 95% of the inactive parameters. The system's complexity is further highlighted by the repository, where 95.2% of the code consists of Lean 4 mathematical proofs, and the entire 114-page paper and code were largely generated by Claude. Human oversight is limited to finalizing six critical security definitions: completeness (false rejection rate below one in a trillion), knowledge soundness (forgery probability below one in 2 to the 100th power), and zero-knowledge. In this framework, humans establish the axioms, while AI performs the reasoning, leading to a safety contract for advanced AI development.