Anthropic published a blog post on Tuesday describing FlashProve, an open-source system designed to cryptographically prove which AI model produced a given output. The project aims to provide what researchers call “provable inference” — a mathematical assurance that a particular answer originated from a specified model.
Purpose and use cases
FlashProve is analogous to cryptographic email signatures: it would allow recipients to confirm that a message truly came from the claimed sender. The system is intended to be robust against impostors attempting to present one model as another, including attempts by the lab that built and operates a model to disguise which model generated an output. The researchers note this could prevent an AI company from secretly substituting a cheaper model to cut costs or swapping in a more expensive model to inflate reported performance.
The mechanism could also validate third-party claims about which model performed a task. For example, a developer might assert that their software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company might claim that its business plan was reviewed by an appropriate model. FlashProve is intended to provide a way for users to verify those claims.
How it works
For FlashProve to function, AI labs must publish a “fingerprint” for each model they release; that fingerprint can later be matched to specific outputs and cryptographically verified. Crucially, these fingerprints do not reveal a model’s weights — the internal numerical parameters that determine behavior and which closed-model providers such as Anthropic and OpenAI keep private. This design allows labs to publish fingerprints without exposing proprietary weights, making the system compatible with commercial models.
The system is also designed to be lightweight so that verification could be done on a local device.
Open source release and current status
Alongside the announcement, the researchers released a detailed technical specification and working code components on GitHub under an open-source license. They emphasize that research on the system is ongoing and that there are no plans to deploy FlashProve at Anthropic or any other lab at this time.
“This research post is meant to share intermediate progress, and not a completed research agenda,” the researchers write, adding that they are sharing the work so others can review, adopt, and improve it.
Why this matters
By enabling cryptographic verification of which model produced an output without disclosing model weights, FlashProve could increase transparency and accountability for services built on large models. That could help users and third parties gain greater confidence in claims about which models were used for specific tasks while preserving providers’ proprietary assets.
