Anthropic is developing FlashProve, an open source system designed to cryptographically verify which AI model produced a specific output. The system works similarly to email signatures, providing a mathematical guarantee that an answer came from a particular model rather than an impostor.
FlashProve addresses risks including AI companies secretly swapping in cheaper models to cut costs or using more expensive ones to inflate performance results. The system could also verify third-party claims that a task was completed by a specific advanced model, such as a vulnerability scan performed by a frontier cybersecurity model or a business plan reviewed by an appropriate model.
The system requires AI labs to publish a "fingerprint" for each released model that can be cryptographically matched to outputs. Crucially, these fingerprints do not reveal a model's weights, the internal parameters that define its behavior and which closed-model makers like Anthropic and OpenAI keep private. This allows labs to publish fingerprints without disclosing their most valuable assets, enabling the system to work with commercial models. Verification is designed to run on local devices.
Anthropic released technical specifications and working code publicly on GitHub under an open source license. Researchers note that this announcement shares intermediate progress on an ongoing research agenda, not a completed one. The lab states there are currently no plans to deploy the system at Anthropic or any other lab.