Anthropic is developing an open-source framework aimed at verifying the AI model responsible for producing a specific output, according to hinews. This initiative, named FlashProve, is designed to deliver what researchers refer to as “provable inference”—a cryptographic proof that serves as a mathematical assurance that a particular output originated from a designated model. The purpose of this system is to enable users to validate that the response they received indeed came from the model they believe they are utilizing.
Similar to cryptographic email signatures that allow recipients to confirm the authenticity of a message’s sender, FlashProve is engineered to withstand attempts by impostors to misrepresent one model as another, even if the impostor is the very lab that developed and operates the model. The researchers emphasize that this system could prevent an AI company from covertly substituting a less expensive model to reduce expenses or a more costly model to exaggerate performance metrics. Additionally, FlashProve could serve to authenticate assertions made by third parties regarding the execution of specific tasks by an advanced model. “A developer might assert that their software library was evaluated for vulnerabilities by a cutting-edge cybersecurity model, or a company might claim that its business plan was approved by an appropriate model,” the researchers state. “How can a user verify these claims?”
For the system to function effectively, AI laboratories must publish a “fingerprint” for each model they release, which can subsequently be matched to specific outputs and verified through cryptographic means. Importantly, these fingerprints do not disclose a model’s weights—the internal parameters that dictate its behavior and which proprietary model creators like Anthropic and OpenAI keep confidential. This allows labs to share fingerprints without compromising their most critical asset, thereby enabling the system to be compatible with commercial models as well. Furthermore, the system is designed to be lightweight, permitting verification to be conducted on local devices.
In conjunction with the announcement, researchers provided a comprehensive technical specification and operational code for various components of the system, which is accessible under an open-source license on GitHub. Research on the system is ongoing, and there are currently no intentions to implement it at Anthropic or any other laboratory. “This research post is intended to communicate intermediate progress, rather than a finalized research agenda,” the researchers clarify in the announcement. “We are sharing this explicitly with the intention that others examine our work, build upon it, and enhance it.”
