Anthropic's new prototype proves which model wrote an AI response without exposing its weights | AI Archive

Anthropic's new prototype proves which model wrote an AI response without exposing its weights

decoder · 2026-11-04T15:23:49+00:00

Anthropic has released FlashProve, an open-source research prototype that utilizes zkSNARK cryptography to verify whether an AI response was generated by a specific claimed model. This system allows providers to generate a compact digital proof of the model's involvement without exposing its underlying weights or architecture. While current limitations restrict proof generation to contexts of up to 1,024 tokens and deterministic sampling, the technology represents a significant efficiency leap over previous methods. Proof generation for large models occurs at approximately 110 tokens per second on an H100 GPU, with verification taking just 17 seconds on a standard CPU. Anthropic envisions use cases for regulators checking approved models, companies proving safety reviews, and detecting server-side tampering. The project fulfills a goal from Anthropic's Frontier Safety Roadmap to create inference verification tools without relying on trusted hardware. Full code and documentation are available on GitHub under the Apache 2.0 license.