Anthropic releases FlashProve to verify AI outputs using cryptography
The open-source zero-knowledge system proves an authentic model generated a response while keeping proprietary neural weights secret.
Anthropic has released FlashProve, an open-source cryptographic system that verifies whether an artificial intelligence model generated a specific response without exposing its proprietary weights. The tool allows customers, auditors and regulators to confirm that a provider ran the exact model promised rather than a cheaper substitute, addressing a core intellectual property challenge in cloud computing.
Mathematical proofs replace blind trust
Developers increasingly access advanced machine models through third-party APIs and routing services. Today, users have no technical mechanism to confirm whether a provider truly ran the requested system or swapped in a smaller, lower-cost alternative to reduce operational expenses. Interesting Engineering reports that Anthropic designed FlashProve to solve this authenticity dilemma without forcing model owners to reveal proprietary parameters to the public.
The system relies on zkSNARK, a zero-knowledge cryptographic proof method that confirms a calculation happened correctly while shielding underlying values. FlashProve executes in two stages. First, the provider publishes a cryptographic commitment representing a unique digital fingerprint of the neural network weights. Second, when a query arrives, the system outputs the text alongside a mathematical proof tied to that initial commitment. Anyone possessing the prompt, output and commitment can verify the proof on ordinary hardware.
Benchmark performance on a 400-billion parameter model
Anthropic tested the software against Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model. Testing on a single NVIDIA H100 GPU, FlashProve generated cryptographic proofs at a rate of approximately 110 tokens per second according to Interesting Engineering.
Each generated proof occupies around 19 MB for a 1,024-token prompt. Verifying that proof requires no specialized accelerators. A standard 24-core central processing unit completes the verification calculation in roughly 17 seconds, allowing client-side validation on ordinary computing equipment.
Technical overhead and open source release
The prototype currently faces steep computational hurdles before commercial deployment. Generating FlashProve proofs demands 70 to 95 times the GPU run time of standard inference engines such as vLLM as noted by Interesting Engineering. The software handles context windows up to 1,024 tokens, relies on greedy decoding and evaluates integer representations rather than full floating-point operations.
While the code masks specific parameter weights, it leaves the overarching architecture visible. To spur further technical innovation, Anthropic published FlashProve under the Apache 2.0 licence, releasing the executable code, protocol specification, formal verification proofs and technical report for independent inspection.
Read also on The Patent: Transmitting Odors Over the Internet with Scentronic · Penn Engineers Speed Up Event Camera Vision to 440 Times per Second
