Anthropic releases FlashProve to verify AI model origins

Verifying AI model identity

Anthropic has introduced a new open-source system called FlashProve. This tool is designed to cryptographically verify that the AI model a user selects is the same one that actually generated a specific response. It aims to solve the problem of opaque practices in AI services where providers might secretly change the underlying technology.

For example, a service might claim to use a high-end model but secretly route requests to a cheaper, less capable version to save on costs. FlashProve targets these types of hidden maneuvers by providing a mathematical guarantee of the model's origin. It allows for a level of transparency that was previously difficult to achieve without exposing sensitive data.

As NewsCord reported, Anthropic researchers published the technical specifications and some working code for this system on GitHub. They described the core approach as 'Provable Inference.' This means the system creates a proof that a specific model produced a specific answer, allowing for independent verification by the end user.

How zero-knowledge proofs work

The system relies on a family of cryptography known as zero-knowledge proofs. These are mathematical methods that allow one party to prove to another that a statement is true without revealing any extra information. In this context, it means a user can confirm which model was used without ever seeing the model's internal weights.

Model weights are the internal numbers and parameters that define how an AI functions. Because these weights are often proprietary and highly sensitive, companies are usually unwilling to share them. FlashProve allows the provider to send a short proof along with the AI response, which the user can then quickly check on their own device.

Specifically, FlashProve uses a cryptographic object called a 'zero-knowledge succinct argument.' This allows the verification process to be processed quickly. The provider sends the response and a short proof, and the user can confirm the model's identity without needing to know the model's architecture or its internal numbers.

Performance and technical limits

While the system provides high security, it currently comes with a significant computational cost. Anthropic noted that generating these proofs requires much more power than simply running the AI model itself. For instance, the work described as intermediate progress requires about 70-95x the GPU time of running the model on a standard platform like vLLM.

To give a sense of scale, a single H100 GPU can generate proofs at a rate of around 110 tokens per second for a specific model. For a request of 1,024 tokens, the resulting proof size is approximately 19 MB. These numbers help developers understand the trade-offs between high-level verification and the speed of the AI interaction.

The researchers have released several components to help the community test the technology. This includes a fast verifier written in C++ and a provably correct verifier written in Lean. Lean is a formal verification language used to ensure that software and mathematical proofs are logically sound and free of errors.

Current status and future use

Anthropic has clarified that FlashProve is still in an early research stage. They stated that they are sharing interim progress rather than a completed research result. Because of this, the system does not yet offer a full suite of guarantees for every type of AI interaction or executable code.

The technology could eventually be used in a wide range of applications, from consumer products to internal corporate security. It would be particularly useful for independent evaluators who need to verify that a company is serving the specific model they claim to be using. It creates a standard for accountability in the AI industry.

However, the company warned that the system does not automatically extend guarantees to quickly executable code. They also noted that it does not ensure that the code, the report, and the formal version fully match. For now, it serves as a foundational step toward making AI model usage more verifiable and transparent for everyone.

Licensing and accessibility

By releasing the code under the Apache 2.0 license, Anthropic is making the technology available for others to build upon. This open-source approach allows developers to integrate the verification system into their own tools or services. It encourages a standard where model origin can be checked by anyone with the right software.

The goal of the project is to provide a way for users with limited computing resources to verify responses. Even if a user doesn't have the power to run a massive AI model themselves, they can still run the small verification check. This levels the playing field between large AI providers and the people who consume their services.

As the research matures, the community will likely see more optimizations to reduce the GPU time required for proof generation. Reducing that overhead is key to making FlashProve a practical tool for real-time applications. For now, it stands as a significant step toward cryptographically securing the identity of AI models in a crowded marketplace.