Anthropic has announced its development of an open-source platform designed to authenticate the source of specific outputs generated by AI models. This initiative, named FlashProve, aims to deliver what researchers describe as “provable inference” — a type of cryptographic validation that confirms a particular output originated from a designated AI model. The goal is to allow users to verify that the results they receive correspond to the model they believe they are utilizing.
Similar to how cryptographic signatures on emails let recipients confirm the legitimacy of the sender, FlashProve is intended to be resistant to attempts by impostors to misrepresent one model as another, including from the developers themselves. The creators emphasize that this technology could thwart the potential for an AI company to secretly replace a higher-cost model with a cheaper alternative to reduce expenses, or to substitute it with a pricier version to artificially boost performance metrics.
Moreover, FlashProve could also function to authenticate assertions made by external parties that an advanced model performed certain tasks. “For instance, a developer might assert that their software was evaluated for vulnerabilities by a leading cybersecurity model, or a company might claim that its business strategy was assessed by an appropriate AI model,” the researchers noted. “What mechanisms exist for users to substantiate these claims?”
The successful operation of FlashProve hinges on AI labs sharing a “fingerprint” for each model they introduce. This fingerprint can be subsequently matched to specific outputs and verified through cryptography. Importantly, these fingerprints do not disclose the model’s internal weights, which determine its behavior and are typically kept confidential by companies like Anthropic and OpenAI. This ensures that labs can publish fingerprints without compromising their proprietary technology, allowing the system to also function with commercial models. Moreover, the framework is designed to be lightweight, enabling local device verification.
In conjunction with this announcement, the researchers have released a comprehensive technical outline and operational code for various system components, which are available under an open-source license. Although development on this system is active, there are currently no plans to implement it at Anthropic or any other organization.
“This research sharing is intended to present our progress rather than a finished project,” the researchers noted in their announcement. “We encourage others to evaluate our work, build upon it, and enhance it further.”
