US to Evaluate AI Cyber Risks Pre-Release, Excludes Open-Weight Models

The United States is nearing completion of a new system to assess the cybersecurity risks of cutting-edge artificial intelligence models before they are launched. However, the plan notably excludes open-weight models developed by U.S. companies.

“

The U.S. government will review advanced artificial intelligence models for cybersecurity risks before their release.

This approach means that top-tier models from companies like OpenAI and Anthropic, which do not disclose core information such as model weights, are likely to be included in the review. Similarly, some of Google's Gemini models may also be subject to evaluation. In contrast, open-weight models, where users can access information like model weights, such as Meta's Llama and Nvidia's NeMo, will initially be left out. The Wall Street Journal noted, however, that these open-weight models could be included in future reviews if their performance significantly increases.

The White House discussed the specifics of this plan with representatives from Meta, Anthropic, OpenAI, Google, and Nvidia. This initiative is a concrete step to implement an executive order signed by President Trump in June. The order requires companies to provide the federal government with up to 30 days of access to their models before releasing them to trusted partners. While this process is voluntary, industry insiders believe companies will find it difficult to bypass government review due to potential liability issues if safety concerns arise later.

“

Open-weight models from U.S. companies will be excluded from the initial evaluation.

Critics, as reported by The New York Times, argue that while the guidelines formalize government review of leading companies' models, they delay urgent responses to potential threats from open-weight models, especially those developed in China. The increasing availability of powerful and affordable Chinese AI models has led some executives at OpenAI and Anthropic to question U.S. technology practices and call for related regulations. Conversely, some tech industry figures, including Nvidia CEO Jensen Huang, have advocated for expanding the open-weight AI ecosystem to bolster U.S. competitiveness and cybersecurity.

“

This approach aims to balance innovation with safety, but may leave a gap in verifying potential threats.