Anthropic's new tool verifies identity of AI models
What's the story
Anthropic has announced the development of an open-source system called FlashProve. The goal of this innovative platform is to provide "provable inference" or a cryptographic proof that a particular output was generated by a specific AI model. The technology would help users confirm the authenticity of the responses they receive from their chosen models, much like cryptographic email signatures do for messages.
Security feature
Preventing model impersonation
FlashProve is specifically designed to thwart impersonators trying to pass off one model as another, even if the impostor is the lab that created and operates the model.
The technology could be particularly useful in preventing an AI company from secretly swapping in a cheaper model to cut costs, or a more expensive model to inflate its performance results.
Verification tool
Verifying 3rd-party claims
FlashProve could also be used to verify third-party claims about specific tasks being performed by an advanced model.
For instance, a developer might claim their software library was scanned for vulnerabilities by a frontier cybersecurity model, or a company could say its business plan was approved by an appropriate model.
The system would provide users with a way to verify these claims.
System requirements
How it works
For FlashProve to work, AI labs need to publish a "fingerprint" of each model they release. This fingerprint can later be matched with specific outputs and verified cryptographically.
The good news is that these fingerprints don't reveal a model's weights, the internal numbers that define its behavior.
This way, labs can publish the fingerprints without compromising their most valuable asset, making the system applicable to commercial models as well.
Open-source access
FlashProve is lightweight and open-source
FlashProve is designed to be lightweight, allowing verification on a local device.
As part of the announcement, researchers have shared a detailed technical specification and working code for different parts of the system. This is publicly available under an open-source license on GitHub.
However, it should be noted that research in this area is ongoing, and there are no plans to deploy it at Anthropic or any other lab just yet.