Technology and Scienceupdated 2h ago2 min read
Anthropic Releases FlashProve To Cryptographically Verify AI Model Behind Each Response
14outlets comparedWhere the coverage agrees
8 Other2 Local Western2 Western Alternative2 Other source typesThis article is an automated aggregation of coverage from news outlets. Our editorials offer in-depth analysis curated by the NewsCord team.
Anthropic unveiled FlashProve, an open-source system using cryptography to verify model origin. Aims to prevent intermediary model swapping by cryptographic proofs while hiding weights.
Save coverage image
FlashProve model verification
Anthropic released FlashProve, an open-source system that cryptographically verifies whether the AI model a user selected matches the model that actually generated a response. Anthropic researchers said they published for free on GitHub the technical specification and some working code for the system, named "FlashProve," and they described the approach as "Provable Inference." FlashProve uses cryptography from the zero-knowledge proof family so verification can be processed quickly on a user's own device without disclosing the model's weights.
FlashProve targets opaque practices in AI services, including cases where routing secretly uses a cheaper model or deploys a more capable model than advertised. Anthropic said the system is still in an early research stage and that it is "sharing interim progress rather than a completed research result."
How it works
FlashProve is based on "provable inference," a cryptographic proof that indicates that a given AI model generated a response. The goal of "provable inference" is to allow a user with limited computing resources and no information about the weights to verify that a response came from the requested model. The mechanism attaches a short proof to an AI response, so a user can confirm which model produced the answer without revealing the model's internal numbers.
FlashProve uses a cryptographic object called 'zero-knowledge succinct argument' and the provider sends, along with the response, a short proof that the user can quickly verify. FlashProve releases executable code and Lean proofs under the Apache 2.0 license, including a fast verifier in C++ and a provably correct verifier in Lean.
Speed, limits, and stakes
A single H100 GPU generates proofs at around 110 tokens/second for Llama 4 Maverick, with a proof size of ~19 MB per 1,024-token request. FlashProve is about ~70-95x the GPU time of running the model itself on vLLM, and it described the work as intermediate progress rather than a completed research agenda. The commitment to the model and subsequent inference proofs hide the weights of the model, while hiding the architecture of the model is not a goal of this system.
The technology could enable applications ranging from consumer products to internal security, including cases where independent evaluators need to know the model being served. Anthropic warned that the system does not automatically extend guarantees to quickly executable code, and it said "it does not ensure that the code, the report, and the formal version fully match."