Anthropic Releases FlashProve to Verify AI Model Identity Using Cryptographic Proofs

Anthropic Introduces Open-Source System for AI Model Verification

Anthropic has unveiled FlashProve, an open-source cryptographic system designed to verify AI model identity and confirm that outputs genuinely originated from specific models1

1

2

2

. The system addresses a critical trust gap in AI deployment by providing what researchers call "provable inference"—a mathematical guarantee that a given response came from the exact model users believe they're accessing. Similar to cryptographic email signatures that verify sender authenticity, FlashProve creates verifiable links between models and their outputs, making it substantially harder for providers to secretly swap models or tamper with responses.

How FlashProve Uses Zero-Knowledge Proofs to Verify Model Outputs

FlashProve employs zero-knowledge proofs, specifically a cryptographic technique called zkSNARKs (zero-knowledge succinct non-interactive argument of knowledge), to verify model outputs without exposing proprietary weights2

2

. The system operates in two stages: first, AI labs publish a cryptographic "fingerprint" or commitment to their model weights, creating a unique identifier without revealing the internal numbers that define how models behave1

1

. When users submit requests, providers generate cryptographic proofs demonstrating that the committed model produced the corresponding output tokens. Anyone with the commitment, prompt, and response can verify the proof independently, enabling third-party verification of AI systems without requiring access to the underlying model.

Preventing Model Substitution and Ensuring Compliance Verification

The system is specifically designed to prevent model substitution, where providers might secretly swap in cheaper models to reduce costs or more expensive models to inflate performance results1

1

. FlashProve enables users to confirm they're receiving outputs from the advanced models they're paying for, addressing concerns about providers routing requests to alternative systems. Beyond preventing cost-cutting substitutions, the technology supports compliance verification and auditing use cases. Companies could require providers to demonstrate that cybersecurity software was examined by specific frontier models rather than cheaper alternatives2

2

. Regulators could verify whether providers are serving approved models or meeting commitments on AI development, creating accountability mechanisms that don't currently exist.

Performance Metrics and Technical Specifications of the Research Prototype

Anthropic demonstrated FlashProve using Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model running on a single NVIDIA H100 GPU2

2

. The research prototype generates proofs at approximately 110 tokens per second, with each proof measuring around 19 MB for a 1,024-token request. Verification takes approximately 17 seconds using a 24-core CPU without requiring GPU resources. However, generating these cryptographic proofs demands substantially more computing power than ordinary inference—measurements show the process requires approximately 70 to 95 times the GPU time needed by vLLM, a widely used open-source inference engine2

2

. The current system supports contexts up to 1,024 tokens and proves an integer version of models under greedy decoding, though it doesn't yet cover every floating-point or sampled inference result.

Open-Source License Release and Future Development Roadmap

Anthropic has released FlashProve's executable code, technical report, protocol specification, and formal verification proofs under the Apache 2.0 open-source license on GitHub1

1

2

2

. Researchers emphasized that the announcement represents intermediate progress rather than a completed research agenda. "This research post is meant to share intermediate progress, and not a completed research agenda," researchers explained, adding that they're "sharing this explicitly with the goal that others look at what we have done, take it, and improve it"1

1

. There are currently no plans to deploy FlashProve at Anthropic or any other lab, positioning it as a foundation for collaborative development. The system hides proprietary weights but not model architecture, and ongoing research aims to address computational overhead while expanding supported inference types. Watch for improvements in proof generation efficiency and broader adoption as researchers build on these technical specifications to make model identity independently verifiable across the AI industry.