Anthropic develops FlashProve, an open-source system that uses cryptography to prove which model an AI output came from
Anthropic is developing an open-source system called FlashProve that uses cryptographic proofs to provide mathematical-level guarantees for a specific output, confirming that it indeed came from a designated model. The system is similar to encrypted email signatures and can defend against passing one model off as another; moreover, the fingerprint does not leak model weights, so it is also suitable for commercial closed-source models. Anthropic has published the technical specification and runnable code and open-sourced them on GitHub, but the research is still ongoing, and there are no plans yet to deploy it at Anthropic or other AI labs.
Why read this
AI score 61
Anthropic has published the technical specification and runnable code for FlashProve, so readers can learn about the approach to using cryptography to prove model provenance and its current progress.
Topics
Body · AI translation
Anthropic is developing the FlashProve open-source system: using cryptography to prove which model an AI output comes from
English text is being prepared.
English text is being prepared.
English text is being prepared.
FlashProve can also be used to verify third-party claims that a certain task was completed by an advanced model. The researchers wrote: "A developer might claim that their software library has been scanned for vulnerabilities by some frontier cybersecurity model; a company might also claim that its business plan has been endorsed by an appropriate model. How are users supposed to verify these claims?"
Crucially, these fingerprints do not leak model weights, that is, the internal parameters that determine model behavior. Closed-source model developers such as Anthropic and OpenAI typically protect these parameters as trade secrets. Therefore, AI labs can publish model fingerprints without giving up their most valuable assets, which also means FlashProve can likewise be applied to commercial closed-source models.
In addition, the system is designed to be lightweight, and users can complete verification directly on their local devices.
As part of this release, Anthropic researchers published detailed technical specifications as well as runnable code for each component of the system. This code has been made public on GitHub under an open-source license.
At present, the research related to FlashProve is still ongoing, and Anthropic has not yet planned to deploy this system in its own or other AI labs.
The researchers explained in the announcement: "This research article is intended to share interim progress rather than present a completed research program. The explicit purpose of making these results public is our hope that others can review our work, build on it, and further improve it."