Anthropic develops system to cryptographically verify AI model identity

The need for provable inference

Anthropic recently published a blog post detailing work on a new system called FlashProve. The primary goal of this project is to provide what researchers describe as "provable inference." This refers to a cryptographic proof, which is essentially a mathematical guarantee, that a specific output was generated by a specific AI model.

Currently, users often have to take a company's word that the AI they are interacting with is the one they paid for or expected to use. FlashProve aims to change this dynamic by allowing for independent verification of a model's identity. This ensures that the answer a user receives truly originated from the intended system.

As reported by TechCrunch, the system is designed to be robust against impostors who might try to disguise one model as another. This protection is significant because it would remain effective even if the lab that built and runs the model attempted to mislead the user. It creates a layer of trust that does not rely solely on the honesty of the service provider.

Preventing model swapping

One of the primary use cases for FlashProve is preventing AI companies from secretly swapping models to manage their internal costs. A company might want to use a cheaper, less capable model for certain tasks while charging the user for a more expensive, high-performance model. FlashProve would make it mathematically impossible to perform this swap without the user noticing.

Conversely, the system could prevent a company from inflating its performance results by using a more expensive model than advertised. By providing a verifiable link between the output and the specific model, the system ensures that the capabilities being sold are the ones actually being delivered. This creates a level of accountability for how companies deploy their technology.

The researchers also suggest that this could be used to verify claims made by third parties. For example, a developer might claim their software was scanned for vulnerabilities by a frontier cybersecurity model. With FlashProve, a user could verify that the scan was actually performed by that specific advanced model rather than a simpler, less effective one.

How cryptographic fingerprints work

To make this system function, AI labs would need to publish a "fingerprint" for each model they release. These fingerprints act as unique identifiers that can be matched to specific outputs and verified cryptographically. This allows for a verifiable chain of custody from the model's architecture to the final text or image produced.

A critical technical detail is that these fingerprints do not reveal a model's weights. Weights are the internal numbers that define how a model behaves, and companies like Anthropic and OpenAI keep these private as their most valuable assets. Because the fingerprints are separate from the weights, labs can participate in this verification system without giving up their proprietary secrets.

The design of FlashProve also prioritizes efficiency and accessibility. The researchers designed the system to be lightweight, meaning that the actual verification process could be performed on a local device. This ensures that users do not need massive computing power just to check if an AI output is authentic.

Open source and research goals

Anthropic has released the technical specifications and working code for various elements of the system under an open-source license on GitHub. By making the code public, they are inviting the broader research community to examine, test, and build upon their initial work. This approach allows for faster iteration and wider adoption of the technology.

However, it is important to note that Anthropic does not currently have plans to deploy FlashProve at their own labs or any other existing platforms. The current release is intended as a way to share intermediate progress rather than a finished product. The researchers are providing a foundation for others to develop further.

The team expressed that their goal is to share what they have done so that others can take it and improve it. By providing the code and specifications now, they hope to foster a standard for AI identity verification. This collaborative approach helps establish a framework for transparency before the technology becomes a standard industry requirement.

The path toward verification

The development of FlashProve represents a move toward a more transparent AI ecosystem where model identity is verifiable. As AI models become more integrated into business plans and cybersecurity workflows, the ability to prove which model did the work becomes increasingly important. It addresses a growing concern about the reliability of AI-driven services.

For developers and companies, this could eventually lead to a standard for