Anthropic is researching FlashProve, an open source system that cryptographically verifies the origin of outputs generated by AI models. The system aims to prove, through “provable inference,” that a specific output was generated by a designated model, and has been presented as a new approach to AI model verification. Published technical specifications and some working code are available, but FlashProve remains at an early research stage, with no plans to release it through Anthropic or another research institution.
How FlashProve detects outputs disguised as those from another model
FlashProve is designed to let recipients confirm whether a message actually came from its sender, much like cryptographic email signatures. It is intended to withstand attempts to pass off one model as another and to verify claims that an AI company has substituted a less expensive model to reduce costs or used a more expensive model to inflate performance claims.
The researchers considered cases in which a developer claims that a state-of-the-art cybersecurity model checked a software library for vulnerabilities, or a company claims that an appropriate model reviewed its business plan. FlashProve was proposed as a way to verify which model is actually associated with such claims.
For this purpose, an AI research lab would need to publish a “fingerprint” for each model it releases, link that fingerprint to specific outputs and verify the link cryptographically. Just as an email signature identifies the sender of a message, a model fingerprint provides information that identifies the origin of an output. TechCrunch’s report on FlashProve describes the system as a research project focused on verifying model identity.
Verifying closed models without revealing their weights
A model fingerprint does not expose the model’s internal weights. Weights are the internal numerical values that define how a model operates, and companies that operate closed models generally keep them secret. A research lab could therefore publish a fingerprint without giving up its most important asset, and FlashProve could potentially be applied to commercial models as well.
The verification process is also designed to be lightweight. FlashProve is intended to run verification on a local device, allowing users to check the origin of an output without relying on a separate centralized service. However, it has not been confirmed whether the system works with every closed model in a real service environment, and the current release consists of interim research results.
Potential practical uses and remaining challenges
FlashProve could be used to verify claims that a third party used an advanced model for a specific task or to check the results of software security reviews. However, the system remains a research result with published technical specifications and working code for several components, and no deployment schedule has been announced.
The researchers described the release as an intermediate research step intended to share the work so that other researchers can build on the existing results. FlashProve’s technical specifications and some working code have been released under an open source license, with no plans for a future release.
Source
This article was written using the following source: TechCrunch — Anthropic is working on a system to prove AI models’ identity

Leave a Reply