Anthropic has introduced FlashProve, an open-source research system designed to cryptographically verify that an AI-generated response came from a specific model. The technology uses mathematical proofs to help users and organizations confirm the identity of an AI model without requiring access to its underlying proprietary information.
According to Anthropic’s official research announcement, published on November 2, 2026, FlashProve addresses a limitation in current AI services, where users cannot independently determine whether the model responding to their requests is the one advertised by its provider.
FlashProve Uses Cryptographic Proofs to Verify AI Outputs
FlashProve uses a technique known as zero-knowledge succinct arguments of knowledge, or zkSNARKs, to generate cryptographic evidence that a response was produced by a specified AI model. The process allows users to verify an output without independently running the model or accessing its internal parameters.
AI providers must first publish a cryptographic commitment, effectively a fingerprint that identifies a particular model. Each response can then be accompanied by a proof that allows independent verification against that commitment.
The system is designed to prevent providers or intermediaries from secretly substituting another model, including using a cheaper alternative to reduce operating expenses. It could also prevent companies from using more capable models while presenting evaluation results as though they came from less powerful systems.
Importantly, the verification process does not disclose the model’s weights, which determine how an AI system processes information. This makes the technology potentially suitable for commercial AI providers that keep their models proprietary.
Researchers Demonstrate Verification on Large AI Models
Anthropic’s researchers tested FlashProve using Meta’s Llama 4 Maverick, a mixture-of-experts model containing approximately 400 billion parameters. Their implementation generated proofs at approximately 110 tokens per second using a single Nvidia H100 GPU.
For a request containing 1,024 tokens, the resulting proof measured approximately 19 megabytes. Generating these proofs required roughly 70 to 95 times the GPU computation needed to run the model through conventional inference software.
Despite the computational cost of generating proofs, Anthropic said verification could be performed within seconds using a standard CPU. The researchers also developed mathematical security guarantees using Lean, a formal verification programming language.
Anthropic Publishes Open-Source Research Tools
Anthropic has released a technical specification, executable software, verification tools, and mathematical proofs supporting FlashProve. The materials are available through the company’s GitHub repository under the Apache 2.0 open-source license.
The researchers identified potential applications in cybersecurity audits, AI model evaluations, and verification of claims involving AI-generated work. For example, organizations could confirm that a particular AI security model performed a vulnerability assessment.
Anthropic described FlashProve as an experimental research prototype rather than a completed commercial system. The company has not announced plans to deploy the technology in Claude or other production AI services.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.