Anthropic releases tool to verify which AI model generated a response
The open-source FlashProve prototype uses cryptographic proofs to verify model identity without exposing its weights, but has significant computing costs and limits.
Anthropic has made FlashProve publicly available, releasing an experimental system designed to verify whether an AI response was generated by the model its provider identified. The company published source code, a technical report and formal proofs, while cautioning that the research remains unfinished.
The approach requires providers to publish a fingerprint representing their model’s weights and attach a cryptographic proof to each response. Anthropic said this allows verification without disclosing the weights. Potential uses include checking that model-routing services deliver the model customers paid for and helping independent evaluators confirm which model they are testing.
For testing, Anthropic used Meta’s Llama 4 Maverick, a mixture-of-experts model with 400bn parameters. One Nvidia H100 generated proofs at about 110 tokens per second; each request containing 1,024 tokens required a proof of about 19MB. According to Anthropic, generating proofs consumed 70 to 95 times as much GPU time as ordinary model operation. Verification took about 17 seconds using an ordinary CPU, without requiring either a GPU or the weights.
Anthropic said Claude contributed to development. To address difficulties auditing Claude’s choices, researchers expressed the main security guarantees as proofs checked by machine using Lean, a formal verification language. The software carries an Apache 2.0 license.
FlashProve supports only mixture-of-experts models and contexts limited to 1,024 tokens. It conceals weights but exposes model design, and does not address timing attacks. TechCrunch reported that neither Anthropic nor other labs plan to deploy it. Anthropic said a separate system that signs traffic and checks it again on a trusted server had met its own output-verification goal.
