- FlashProve is an open-source cryptographic system that verifies the authenticity of AI model outputs.
- The system uses zero-knowledge cryptographic techniques to generate mathematical proofs without revealing model weights.
- FlashProve helps detect tampering of model responses and supports independent evaluations.
- Currently a research prototype, it requires more computational resources than standard inference.
Anthropic has announced the launch of FlashProve, an open-source cryptographic system designed to verify whether an artificial intelligence model genuinely generated a specific response without revealing its proprietary weights. This technology helps users, independent auditors, and regulatory bodies confirm whether vendors are operating the models they claim to be using. The system utilizes zero-knowledge cryptographic techniques to generate mathematical proofs that demonstrate a specific model produced a given output token in response to a specific prompt. As companies increasingly rely on third-party model vendors and routing services to access advanced AI systems, this technology aims to address the growing trust issues.
Currently, users cannot reliably determine whether a response comes from the model they selected or if the vendor secretly substituted it with a cheaper alternative. FlashProve aims to provide this assurance without requiring users to access or independently run the model.
How the Cryptographic Technology Works
The cryptographic proof verifies output. FlashProve employs a technique known as zero-knowledge succinct non-interactive arguments of knowledge (zkSNARK), which allows vendors to prove the correctness of computations without revealing sensitive information about the underlying model. The system operates in two phases. First, the vendor publishes a cryptographic commitment of its model weights, effectively creating a fingerprint of the model used while keeping the weights private. When a user submits a request, the vendor can subsequently generate a proof demonstrating that the committed model produced the corresponding output token. Anyone with the commitment, prompt, and response can verify the proof without needing access to the model weights.
Anthropic demonstrated this method using Llama 4 Maverick, an expert mixture model with 400 billion parameters. When run on a single NVIDIA H100 GPU, FlashProve generates proofs at a rate of approximately 110 tokens per second, with each proof being about 19 MB in size for a request of 1,024 tokens. Verification using a 24-core CPU takes about 17 seconds, without the need for a GPU.
Challenges and Limitations of Technical Applications
Security faces practical challenges. The application of this technology is not limited to verifying whether customers receive the models they selected. Anthropic believes that cryptographic inference proofs can help detect tampering of model responses, support independent evaluations, and provide regulators with a way to verify which models vendors are offering. For example, a company could require a vendor to prove that its cybersecurity software was checked by a specific model rather than a cheaper alternative. Regulators could also use such proofs to check whether vendors are providing approved models or meeting certain commitments in AI development.
However, FlashProve remains a research prototype rather than a production-ready solution. The computational resources required to generate proofs far exceed those needed for standard inference. Anthropic's measurements indicate that the process requires approximately 70 to 95 times the GPU time needed by vLLM, a widely used open-source inference engine. The current system also supports a maximum context of 1,024 tokens and proves the integer version of the model under greedy decoding. It has not yet covered all floating-point or sampling inference results. Additionally, it hides model weights but does not conceal the model's architecture.
Significance for Advancing Research and Development
Anthropic has released the executable code, technical report, protocol specifications, and formal verification proofs of the system under the Apache 2.0 license. The company stated that this release aims to encourage other researchers to build upon it and improve the technology. This research represents an early step toward enabling independent verification of model identity and response integrity, potentially enhancing user confidence in the systems behind their applications.
The Impact of FlashProve on Trust in AI Models
As artificial intelligence technology rapidly advances, trust in the authenticity of its model outputs has become a significant issue. The launch of FlashProve provides users and regulatory bodies with a new verification tool that ensures the models used genuinely meet the claimed functionalities without exposing sensitive model information. This not only helps boost user confidence in AI applications but may also promote market demands for transparency from model vendors. Although the technology is still in the research phase, its potential impact could change the industry's approach to verifying AI models.

