Hajra Naz
Anthropic is working on a new system. It is an open-source project. The goal is to verify which AI model produced a given output. The company shared this news in a blog post published on Tuesday.
The system is called FlashProve. It is built to provide what researchers call “provable inference.” This acts as a cryptographic proof. Essentially, it serves as a mathematical guarantee. It ensures that a specific output came from a designated model. It lets users confirm the source of their answers. They can verify that the response really came from the model they expected.
The mechanism resembles cryptographic email signatures. Those signatures let recipients confirm a message came from a real sender. FlashProve functions similarly. It is robust against impostors. It stops bad actors from disguising one model as another. This protection holds even when the impostor is the lab that built and runs the model.
Read More: Anthropic Signals AI Models May Soon Accelerate Their Own Development
Researchers note a key use case. The system prevents AI companies from cheating. It stops them from secretly swapping in a cheaper model to cut costs. It also stops them from swapping in a more expensive model to inflate performance results.
FlashProve has other applications. It can verify claims made by third parties. For example, someone might claim a specific task was handled by an advanced model. “A developer might claim that their software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company might claim that its business plan was given the go-ahead by an appropriate model,” the researchers write. “How can a user verify these claims?”
The system requires a specific setup to work. AI labs must publish a digital “fingerprint” for every model they release. Users can later match those fingerprints to specific outputs. They can then verify the source cryptographically.
Crucially, these fingerprints do not reveal a model’s weights. Model weights are the internal numbers that define how an AI behaves. Closed-model makers like Anthropic and OpenAI keep those weights strictly private. Because the fingerprints hide the weights, labs can publish them safely. They do not have to give up their most valuable asset. This design allows the system to work with commercial models too. Additionally, the system is lightweight. Verification can happen right on a local device.
Read More: Anthropic Blocks Claude From Cloning Video Games
Alongside the announcement, researchers released technical documents. They published a detailed technical specification and working code for various parts of the system. This material is publicly available. It is licensed under an open-source agreement on GitHub.
Research on the system is still ongoing. There are currently no plans to deploy it at Anthropic or any other lab.
“This research post is meant to share intermediate progress, and not a completed research agenda,” researchers explain in the announcement post. “We are sharing this explicitly with the goal that others look at what we have done, take it, and improve it.”
