US government nears deal on voluntary safety standards for frontier AI models

The US government is negotiating voluntary safety standards with AI companies for releasing frontier AI models, aiming for an announcement as soon as next week.

The proposed standards will set benchmarks to evaluate frontier AI models, establish timelines for their release, and clarify domestic and international access. These standards aim to reduce cybersecurity risks and potential misuse by foreign military or intelligence agencies, especially from nations like China and Russia ^{[1, 2, 3, 4]}.

The discussions involve leading AI firms including OpenAI, Anthropic, Google, Amazon, and Microsoft. Google is actively engaged ahead of releasing advanced coding AI models. OpenAI delayed the full public launch of its GPT-5.6 model after a US government request, limiting access to vetted partners only ^{[1, 3, 4]}.

Anthropic’s advanced Fable and Mythos models were subject to interim US export controls on national security grounds. The Commerce Department imposed and then lifted restrictions in June 2024. Anthropic will collaborate with government and industry partners to help develop voluntary safety and assessment frameworks ^{[1, 2, 3, 4]}.

These efforts trace back to a June 2024 executive order signed by former President Donald Trump, directing federal agencies and AI companies to work together on testing and setting voluntary standards before model release ^{[1, 2, 3, 4]}.

The US AI Standards and Innovation Center (CAISI) and the National Security Agency (NSA) are playing key roles in formulating and overseeing the safety standards ^{[2, 3, 4]}. Negotiations cover how long model reviews should last and what threshold defines a model as 'frontier' ^[2].

OpenAI CEO Sam Altman said he hopes to "establish a global framework for AI standards and provide expert, fair analysis of AI capabilities and risks so benefits can be widely shared," emphasizing the need for objective risk assessment ^[2].

The voluntary standards aim to smooth model launches while reducing risks of misuse. Industry and government hope the framework will provide a consistent, trusted process for evaluating and releasing frontier AI models ^[2].

An announcement detailing the voluntary safety standards could come as early as next week, marking a notable step in US efforts to govern advanced AI technologies ^{[1, 5, 2, 3, 4]}.