Anthropic has introduced FlashProve, an open-source cryptographic tool that helps users verify AI model authenticity, addressing concerns over hidden model switching and transparency in AI services.
Anthropic, the company behind Claude AI, has introduced a new open-source research prototype called FlashProve to improve transparency and trust in artificial intelligence systems. The technology is designed to help users confirm whether an AI-generated response actually comes from the specific model requested.
The development addresses growing concerns that AI companies could secretly replace premium models with cheaper alternatives to reduce operating expenses without informing customers.
FlashProve aims to tackle this issue through cryptographic verification, allowing users to establish the authenticity of AI-generated responses without requiring companies to disclose their confidential model architecture or internal parameters.
FlashProve is a verification system that provides mathematical evidence of which AI model generated a particular response.
As AI services become increasingly popular, customers and businesses are becoming more concerned about transparency. Companies paying for access to advanced AI models may want assurance that their requests are being processed by the promised systems rather than less expensive alternatives.
FlashProve is designed to address this challenge by creating a verifiable connection between an AI model and its output.
The technology could eventually help establish greater accountability among AI developers, cloud service providers and businesses that depend on advanced AI systems.
FlashProve uses cryptographic techniques to verify the authenticity of AI-generated responses. Its approach can be compared to a digital signature used to establish the authenticity of an electronic document.
Before making an AI model available, its developer publishes a unique digital fingerprint associated with that model.
When a user submits a query, the AI server returns the generated response along with a cryptographic proof. This evidence can then be checked to confirm whether the output was produced by the specified model.
An important part of this process is a security technique known as a zero-knowledge succinct non-interactive argument of knowledge (zkSNARK).
This method allows verification to take place without revealing sensitive information about the underlying AI model, including its confidential weights.
Model weights are numerical parameters that determine how an AI system processes information and generates responses. These parameters represent valuable intellectual property for AI companies.
By protecting these assets while allowing independent verification, FlashProve could make it easier for commercial AI developers to introduce transparency without exposing proprietary technology.
The development of FlashProve involved collaboration between Anthropic and researchers from several organisations, including the Massachusetts Institute of Technology (MIT), EPFL, StarkWare and Succinct.
The research team tested the technology using Llama 4 Maverick, a large AI model containing approximately 400 billion parameters.
The experiment was intended to demonstrate that cryptographic verification could work even with extremely large AI systems.
According to the reported performance results, generating proof for a request containing 1,024 tokens produces a file of approximately 19 MB. The system achieved a processing speed of around 110 tokens per second using a single Nvidia H100 graphics processing unit.
These results demonstrate the potential of FlashProve, although generating proofs still requires considerable computing resources.
One of the notable features of FlashProve is the difference between generating and verifying cryptographic proofs.
While producing the proof requires substantial processing power on the AI provider's side, checking the proof is comparatively less demanding.
According to the reported testing results, users can verify an AI response using a standard computer CPU in approximately 17 seconds.
This means customers would not necessarily need expensive graphics hardware or specialised computing infrastructure to confirm the identity of an AI model.
The system could therefore make independent AI verification more accessible to developers, businesses and researchers.
To encourage wider research and development, Anthropic has made FlashProve's code and technical documentation publicly available on GitHub.
The project has been released under the Apache 2.0 open-source licence, allowing developers and researchers to examine, study and build upon the technology under its licence terms.
Making the project publicly accessible could encourage other AI companies and independent developers to explore cryptographic verification methods.
It may also help establish new approaches to transparency and accountability as AI-powered services become increasingly important across industries.
FlashProve could have applications beyond checking whether an AI company is using the model it promises. For example, businesses could use cryptographic evidence to demonstrate that particular software underwent security checks using a specified advanced AI model.
Similarly, financial technology companies and startups could potentially use verification records to establish that an important assessment or business decision was processed through a particular AI system.
Such applications could be valuable in sectors where transparency, accountability and reliable verification are essential.
Despite its potential, FlashProve remains an early-stage research prototype rather than a commercially available product. Currently, the system supports conversations limited to 1,024 tokens, restricting its usefulness for longer AI interactions.
Another major challenge is computational cost. Generating cryptographic proofs reportedly requires approximately 72 to 95 times more computing time than conventional AI operations.
This additional processing requirement could make large-scale commercial deployment expensive.
Anthropic has also indicated that it does not currently plan to introduce FlashProve as part of a live commercial product or API.
Further research and performance improvements will therefore be necessary before the technology can become a practical feature in everyday AI applications.
Conclusion
Anthropic's FlashProve represents an important research development in improving transparency and trust in artificial intelligence. By combining cryptographic verification with zero-knowledge proofs, the technology could help users confirm AI model authenticity without exposing proprietary information. However, high computing costs and current technical limitations mean wider commercial adoption may still take time.