Anthropic is developing an open-source verification system called FlashProve, aimed squarely at a pain point: whether the AI answer you receive was actually generated by the model you think you are using. In a Tuesday blog post, the company said the "provable inference" it aims to build uses cryptography to provide a mathematical guarantee for every output, proving that it truly came from a specified model—much like encrypted email signatures help recipients confirm the sender, FlashProve is specifically designed to prevent "passing off model A as model B."
Closing cost-cheating backdoors—and exposing padded scorecards
Researchers point to a long-standing unspoken rule: with it, AI companies will find it hard to secretly swap an expensive model for a cheaper version in the background to save costs; conversely, businesses also cannot secretly disguise a cheap model as an expensive one or artificially inflate benchmark scores. It can also verify third-party claims—for example, a developer claiming that a codebase has been scanned for vulnerabilities by a frontier security model, or a company claiming that its business plan has been endorsed by the relevant model. Users would then finally have a basis for asking, "Why should we trust you?"
Fingerprints Are Public, Weights Stay Secret, and Verification Happens Locally
The key design: AI labs only need to provide a "fingerprint" for each model, then match outputs against it and cryptographically verify their origin; the fingerprint does not reveal the model weights—the internal parameters that determine behavior, which are the lifeblood of closed-source models. Labs can publish fingerprints without handing over core assets, so FlashProve can also work with commercial closed-source models. The system is deliberately lightweight, so users can verify locally. Anthropic has released the technical specification and runnable code (open-sourced on GitHub), but stresses that the research is still ongoing and it does not yet plan to deploy it itself.
