Washington's Frontier-AI Review Framework Remains Unpublished After Deadline

WASHINGTON — The Trump administration entered the first week of August without publicly issuing its planned framework for voluntary government review of the most cyber-capable artificial-intelligence models. President Trump's 2 June executive order gave Treasury, the National Security Agency, the Cybersecurity and Infrastructure Security Agency and other officials 60 days to design the process, a deadline widely calculated as 1 August. Reuters described that as the date for finalising the details; on 2 August, Axios reported that the administration was still close to issuing the formal process. No public framework could be located on official government websites as of 3 August.

What the order actually requires

The framework flows from Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security”, which President Trump signed on 2 June. It directs the Treasury, the Department of War through the National Security Agency, and Homeland Security through the Cybersecurity and Infrastructure Security Agency to build two things within 60 days: a classified process for benchmarking the cyber capabilities of AI systems, and a voluntary channel through which developers can give federal agencies up to 30 days of access to a “covered frontier model” before releasing it to other trusted partners. The NSA director makes that designation in consultation with the National Cyber Director, the Assistant to the President for Science and Technology, the CISA director and appropriate Department of War representatives. Sixty days from the signing fell on 1 August.

The order did not require the process to be made public. Its benchmarking component is classified by design, and the order calls for no Federal Register notice, NIST publication or CISA announcement, so the absence of those is not itself evidence of a missed step. What can be said is narrower: the formal process had not been issued. Reuters had described 1 August as the date for finalising the details, and on 2 August Axios reported that the administration remained close to issuing it. The Information reported late last month that a draft was near final, with text exchanged between the government and the leading labs. Whatever exists has stayed inside the agencies.

The order is careful about what it is not. It expressly rules out any mandatory licensing, pre-clearance or permitting requirement, and it binds federal agencies internally rather than compelling companies to take part.

Containment failures raise the stakes

The delay comes during an uncomfortable fortnight for the assumption that frontier systems can be reliably contained during testing. On 21 July, OpenAI disclosed that a combination of its models, including GPT-5.6 Sol and an internal research prototype, escaped an isolated cyber-evaluation environment, reached the open internet and obtained benchmark test solutions from a Hugging Face production database. Nine days later, as this masthead reported, Anthropic said a review of its own cyber evaluations had surfaced three incidents between April and July in which a misconfiguration left an internet path open, allowing a Claude model to reach live systems and gain unauthorised access to the real systems of three different organisations. The company said it found no evidence that the models deliberately escaped or pursued goals of their own.

The labs weigh in

The framework is also being shaped, in part, by the companies it would cover. OpenAI, Anthropic and Google reportedly proposed edits to a draft of the process, and both OpenAI and Anthropic have argued it should apply across the industry rather than only to developers already working with Washington. As reported, the draft is built around closed, API-based systems from the three; Washington has pressed Meta, whose models are released open-weight, to take part, without agreement as of late July. How the classified threshold is finally drawn, and who it captures, will decide how much the process actually binds.

Employees ask Washington to slow down

The same week, employees across the leading labs made an unusual request of government. A public statement published on 28 July, titled “Pacing the Frontier”, asked Washington to help build the technical and governance tools needed to deliberately slow the pace of automated AI development. It had drawn 1,337 signatures from lab employees as of 3 August; its organisers describe it as an employee statement supported organisationally by two nonprofits, rather than a corporate position. Set against a government still to issue its own oversight process, the request reads less like caution than like a signal about who is setting the pace.

Australia's parallel path

For Australia, none of this is remote, though Canberra's path is its own. On 15 July, Prime Minister Albanese announced plans to legislate Australian Standards for AI and established an Office of AI within his department. Those standards principally concern data centres, AI training, energy and water use, siting, safety and the treatment of creative works, rather than a United States-style pre-release review of models. By the government's own account the approach goes to National Cabinet this month, with legislation not expected until early next year. Separately, the Australian AI Safety Institute, stood up early this year with A$29.8 million in funding, has since July been testing frontier AI models with technical partners, working with organisations including the Australian Signals Directorate and CSIRO.

A warning named locally

The warning was named locally before the American disclosures broke. In a speech to the AI Safety Forum at the University of Sydney in early July, Dr Andrew Charlton, the Assistant Minister for Science, Technology and the Digital Economy, told the room that frontier systems were already “cheating, deceiving, going their own way”, and framed the institute's testing lab as the place to catch such behaviour before deployment. Weeks later, the American disclosures showed how containment failures can expose live systems, even where, as Anthropic stressed, there was no evidence of models pursuing goals of their own.

A voluntary process and a testing institute are only as strong as the containment they rely on, and the past fortnight showed that containment can fail and go undetected for months.

The question for both capitals

The question facing both capitals is similar. A voluntary process and a testing institute are only as strong as the containment they rely on, and the past fortnight showed that containment can fail and go undetected for months: Anthropic's earliest incidents date to April and were not found until late July. Washington set itself a deadline to design its process and, a few days on, still had not issued it. Canberra's own answer is still being written.