Anthropic unveils FlashProve to cryptographically verify AI model identity

Anthropic has announced FlashProve, an open-source research prototype designed to provide “provable inference.” The system uses zero-knowledge cryptography, specifically zkSNARKs, to provide a mathematical guarantee that a specific AI model generated a particular output.

FlashProve functions similarly to a digital signature for emails. It allows users, auditors, and regulators to verify the authenticity of an AI response without requiring access to the model’s proprietary weights. To enable this, AI labs publish a cryptographic “fingerprint” or commitment of the model. This allows for verification while protecting the commercial secrets of companies like Anthropic and OpenAI.

The system aims to prevent several issues, including model impersonation by attackers and the practice of AI providers secretly swapping advanced models for cheaper, less capable ones to reduce costs. It can also be used to verify third-party claims, such as whether a software library was actually scanned by a high-end cybersecurity model.

In technical demonstrations using a 400-billion-parameter Llama 4 Maverick model, FlashProve generated proofs at approximately 110 tokens per second on a single NVIDIA H100 GPU. Verification is designed to be lightweight, capable of being performed on a local device using a standard CPU.