Anthropic Builds FlashProve to Prove Which Model Produced an Output
Anthropic has developed FlashProve, a system that lets a service provider prove an output came from a specific AI model.
1 source 6d ago
Reporting from 1 source: GIGAZINE.
Why it matters
Cheap, weight-hiding proof that a given model produced a given output would close the gap that lets a provider advertise one model and serve another.
Key Facts
- Anthropic developed FlashProve to prove that an output was generated by a specific AI model.
- FlashProve uses the zkSNARK cryptographic proof method, which does not include information about model weights.
- Verification of a FlashProve proof file completes in a few seconds on a 24-core CPU.
- Anthropic published the FlashProve code on GitHub as a research prototype that is not maintained.
- A FlashProve technical report is published as a PDF.
It uses zkSNARK proofs that can be generated more cheaply than with existing methods and that omit model weight information, so closed models can be covered without exposing them. Verification takes a few seconds on a 24-core CPU. Anthropic published the code as an unmaintained research prototype on GitHub, with a technical report.
The stated problem is a mismatch between the model a service advertises and the model it runs. Anthropic lists three ways that gap gets exploited: a provider quietly serves a cheaper model for the same price, a provider swaps in a different model after an auditor has cleared the original, or an attacker impersonates a model to reach users. FlashProve is Anthropic's answer, built on zkSNARK proofs. The proof file carries no information about the model's weights, so a provider can hand over proof for a closed model without revealing what is inside it. Verification finishes in seconds on a 24-core CPU. The published repository is labeled a research prototype and is not maintained.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google