Anthropic develops open-source tool for AI model authentication

Drawing a parallel to cryptographic email signatures, FlashProve is intended to prevent fraudulent impersonation of AI models, even by the organizations that created them. This would guard against practices like substituting a less expensive model to reduce costs or using a more powerful model to artificially inflate performance metrics.

The tool could also be used to validate claims made by external entities regarding the involvement of advanced AI models in specific tasks. For example, verifying that a software library was scanned for vulnerabilities by a sophisticated cybersecurity model or confirming a business plan received approval from a suitable model.

FlashProve relies on AI labs publishing a unique “fingerprint” for each model they release. These fingerprints can be matched to outputs and cryptographically verified. Importantly, these fingerprints do not expose a model’s internal workings or weights, protecting proprietary information. This allows the system to be compatible with commercial models.

The system is designed to be efficient, enabling verification on personal devices. Anthropic has made the technical specifications and code available on GitHub under an open-source license.

Researchers emphasize that this is an ongoing project and not a finalized product. They encourage others to review, utilize, and enhance the system, stating, “This research post is meant to share intermediate progress, and not a completed research agenda.”