Anthropic is developing FlashProve, a cryptographic system to prove which AI model generated a response

Anthropic is working on an open-source system to verify which AI model produced a given output, according to a blog post published Tuesday and reported by TechCrunch. The system is called FlashProve and aims to deliver what researchers call “provable inference”: a cryptographic proof, essentially a mathematical guarantee, that a response came from a specific model. The idea is that a user can confirm that the response they received really comes from the model they believe they are using.

The analogy offered is that of cryptographic email signatures, which let the recipient check that a message comes from the stated sender. According to the article, FlashProve would be robust against impostors trying to pass off one model as another, even when the impostor is the very lab that built and operates the model.

Among the uses the researchers propose is preventing an AI company from secretly swapping a model for a cheaper one to cut costs, or for a more expensive one to inflate its performance results. It would also help verify third-party claims about which model performed a task. The researchers give as examples a developer saying their software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company claiming its business plan was approved by a suitable model, and ask: “How can a user verify these claims?”

To work, the system requires labs to publish a “fingerprint” of each model they release, which can then be matched against specific outputs and cryptographically verified. A key point is that these fingerprints do not reveal the model's weights, the internal numbers that define its behavior and that closed-model makers such as Anthropic and OpenAI keep private. Labs could thus publish fingerprints without giving up their most valuable asset, which would allow the system to be applied to commercial models as well. It is also designed to be lightweight, so that verification can be performed on a local device.

Alongside the announcement, the researchers released a detailed technical specification and working code for several components of the system, publicly available on GitHub under an open-source license.

The actual scope is worth bearing in mind: the research is still ongoing and, according to the article, there are no plans to deploy the system at Anthropic or any other lab. The researchers themselves explain that the publication aims to “share intermediate progress, not a completed research agenda,” and that they are sharing it explicitly so that others can “pick it up and improve on it.”