Published 2 days ago • loading... • Updated 7 hours ago
Anthropic Unveils Cryptographic Tool to Verify Model Outputs
Researchers released open-source code and a technical spec for FlashProve, which could let users verify model provenance on a local device.
On Tuesday, Anthropic announced FlashProve, an open-source system designed to provide "provable inference"—a cryptographic guarantee that a specific model produced a given output.
To function, FlashProve requires AI labs to publish a model "fingerprint" that can be matched to outputs and verified cryptographically, without revealing proprietary model weights.
Operating similarly to cryptographic email signatures, the system prevents companies from secretly swapping models to cut costs or inflate performance results through impostor claims.
Researchers released detailed technical specifications and working code on GitHub, explicitly inviting the community to examine, take, and improve the system's various elements.
While this intermediate research is ongoing, Anthropic has no immediate plans to deploy the system, focusing instead on gathering community feedback and further development.