Anthropic unveils method to verify which AI model produced a response

New Anthropic research: progress on provable inference for large language models. Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so. These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers. For our full write-up, see here: Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: We’re grateful to our collaborators:…

Last attempt at proving which model wrote a response: 14 hours and a 93 GB proof. Anthropic’s new one: a few seconds and 19 MB. It’s called FlashProve. Right now you can’t be sure which model answered your prompt. A provider could quietly swap in a cheaper one, or a hacker could change the answer on the way to you. FlashProve adds a proof to each response that you can check yourself, and the model’s weights stay secret. – It runs on a single H100 at 110 tokens per second – It works on Llama 4 Maverick, a 400B model – Checking a proof takes about 17 seconds on a normal CPU – It costs roughly 70 to 95 times the GPU time of just running the model Claude also helped build the improvements. FlashProve is open source now.