Anthropic releases FlashProve, open-source code that cryptographically verifies an AI response came from a given model, not a cheaper one swapped in by a router

You can now verify that the AI you are interacting with is the one you expect, not a cheaper substitute.

Anthropic has released FlashProve, an open-source system designed to cryptographically verify that an AI model's response originated from a specific, claimed model. This addresses the problem where users or evaluators cannot be certain that the AI they are interacting with is the one they intended, as routers or attackers could substitute cheaper or different models. FlashProve uses zkSNARKs, a type of cryptographic proof, to generate verifiable certificates for AI inferences. The system achieves significant speed improvements over prior open-source methods, with a 400B-parameter model generating proofs at approximately 110 tokens per second on a single H100 GPU. Anthropic is releasing the code, technical reports, and formal verification proofs under the Apache 2.0 license, encouraging community development and auditing.

The release of FlashProve marks a significant step towards verifiable AI inference, directly addressing concerns about model substitution and integrity in AI deployments. By leveraging zkSNARKs, Anthropic's system offers a practical method for proving the origin of AI responses without revealing proprietary model weights. This technology has broad implications for AI regulation, auditing, and consumer trust, enabling independent verification of AI behavior and compliance with certain mandates.

This advancement could fundamentally alter how AI services are contracted and monitored, potentially leading to new service level agreements that incorporate cryptographic verification of model identity. The open-source release also invites broader adoption and innovation in provable AI systems, suggesting a future where AI outputs are as verifiable as other forms of digital evidence.