US Reportedly Excludes Open-Weight AI Models From Cyber Tests

Open-weight AI models will reportedly remain outside a new US government program designed to test whether advanced systems can conduct sophisticated cyber operations.

Trump administration advisers communicated the decision during an August 4 White House meeting attended by representatives from Meta, Google, Nvidia, OpenAI, and Anthropic, according to Reuters.

Federal program will focus on proprietary AI models

The voluntary testing framework follows a June 2 executive order that directed federal agencies to create classified benchmarks for advanced AI cybersecurity capabilities.

Officials want to determine whether frontier models can perform complex cyber operations that could threaten government systems, critical infrastructure, or private organizations.

Previous discussions reportedly finalized arrangements for testing powerful proprietary systems, including Anthropic’s Mythos and Fable models and OpenAI’s GPT-5.6 Sol.

However, the current framework will reportedly exclude open-weight models. These systems allow developers to download, inspect, modify, and deploy their underlying model weights instead of accessing them only through a company-controlled service.

Open-weight exclusion could leave a security gap

The reported exemption could create a significant gap in the federal government’s evaluation system.

Open-weight models can spread rapidly because individuals and organizations can copy and modify them without relying on the original developer’s infrastructure. Developers can also remove built-in safeguards, fine-tune models for specialized tasks, or run them privately.

A highly capable open-weight model could therefore reach a large number of users without facing the same federal cybersecurity tests as proprietary frontier systems.

The government has not clarified whether the exclusion covers every open-weight release or only models that fall below a particular capability threshold.

Anthropic recently said it does not support banning open-weight AI models, despite repeatedly warning that advanced AI systems could create serious cybersecurity and national security risks.

Key details about the program remain unclear

The White House has not identified which federal agency will operate the testing program or manage access to the classified benchmarks.

Officials have also not explained whether they will publish any evaluation results. Keeping the findings classified could help protect sensitive testing methods, but it would give researchers and the public little information about the capabilities uncovered during assessments.

The administration has not disclosed what consequences a company could face if its model demonstrates dangerous cyber capabilities. The voluntary nature of the program also raises questions about whether developers could decline additional testing or continue releasing a model after concerning results.

Officials could eventually create a separate framework for open-weight systems, although no such program has been announced.

The decision comes as governments and AI companies examine how autonomous agents interact with real systems. In a separate case, OpenAI and Anthropic agents targeted real people during cybersecurity tests, highlighting the risks that can emerge when advanced models operate beyond controlled environments.