Anthropic creates a system to verify the authenticity of AI models

Anthropic unveiled FlashProve — a system that allows checking whether a response to a user was actually provided by the claimed artificial intelligence model. 

The company said this in a statement.

The technology uses cryptography to prevent unnoticed substitution of models, including with cheaper versions.

How FlashProve works

Today, users of AI services cannot reliably verify which specific model is answering their requests. For example, a company might quietly replace a more expensive model with a cheaper one to reduce costs. Detecting such a substitution based only on response quality is difficult.

FlashProve aims to solve this problem through cryptographic attestation. For this, the AI developer publishes a kind of digital fingerprint of the model that corresponds to a specific set of its parameters.

When a user sends a request, the system can create a mathematical proof that the response was generated by that exact model. It can be verified without access to the model’s internal parameters.

Overall, Anthropic is considering several possible use cases for FlashProve. In particular, the system will allow customers to verify whether they are receiving responses from the model they paid for. The technology can also help detect cases where attackers substitute AI responses or where developers use different models than those claimed.

Another application is verifying work that was supposedly performed by an AI. For example, a software developer could confirm that a specific AI model actually checked their code for vulnerabilities.

Additionally, FlashProve could potentially help independent researchers and regulators verify which models companies are using.

Development stage

Anthropic has already tested FlashProve on Meta’s Llama 4 Maverick model, which has 400 billion parameters. During trials, the system produced proofs at a speed of about 110 tokens per second using a single NVIDIA H100 GPU.

The size of a single proof was approximately 19 MB, and its verification took between 17 and 21 seconds on a CPU depending on the number of computation threads.

At the same time, creating such proofs currently requires significant resources. According to Anthropic, it uses approximately 70–95 times more GPU time than ordinary processing of similar requests.

Anthropic emphasizes that research is ongoing and FlashProve is not ready for wide deployment. The company published interim results so other researchers can audit the technology and improve it.

Recall that Anthropic filed documents for an initial public offering (IPO). The company plans to list on the U.S. Nasdaq exchange, where its shares will trade under the symbol ANT. The developer of Claude filed an S-1 registration statement with the U.S. Securities and Exchange Commission (SEC).

Also read: Anthropic and British investment firm Hg will create a company in Europe to implement AI in business