Trump AI order revives early model reviews for OpenAI

President Donald Trump has signed an artificial-intelligence order that asks frontier model developers to give US officials early access to some powerful systems before public release. In the White House order, the programme is framed as voluntary rather than a licensing scheme for companies such as OpenAI, Google and Anthropic.

Developers may share models with the US government for up to 30 days before launch, according to The Guardian. That window is narrower than the tougher version discussed earlier, but it brings back pre-release scrutiny after the administration had signalled a lighter approach.

In practice, the shift does not create an approval gate. It gives officials a route into classified testing, cyber-risk checks and private feedback before a major release. Australian enterprises that use US-built AI products and APIs will mainly see the policy as another sign that model safety is moving into national-security channels, not staying inside vendor trust pages and benchmark tables.

The text tells officials to build a process for cyber testing of advanced models:

develop and maintain a classified benchmarking process to assess the advanced cyber capabilities of AI models
Source: White House order

Agencies also have 60 days to create an AI cybersecurity clearinghouse. That body is intended to share information about model capabilities, vulnerabilities and defensive uses across the US government. Such a system would let Washington learn from pre-release testing without making every model launch conditional on a formal permit.

Axios reported that an earlier version of the policy considered a 90-day review period. The final order cuts the window to 30 days while keeping the premise that the government should see some frontier models before customers do.

Business Insider described the order as a scaled-back version of a proposal that drew attention during debate over frontier-model launches, including Anthropic’s Mythos release. Routine chatbots are not the concern. The order is aimed at the small group of systems that may show advanced cyber, coding or autonomous-agent capability before independent testers and customers have had a look.

Voluntary review, not licensing

One carve-out matters more than the rest. The early-access section cannot be used to create mandatory preclearance, the order says:

Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement
Source: White House order

That wording separates the Trump administration’s approach from a full licensing model. Model developers can participate, but agencies do not gain a new power to block a release. The White House can claim a security role without creating a hard approval system for labs.

For OpenAI, Google and Anthropic, the operational question is whether voluntary access becomes part of the launch calendar. A company preparing a frontier release could give officials a confidential look, take feedback on dangerous cyber capability, then still publish on its own timetable. Skipping the process would not breach the order, based on the text released by the White House.

That leaves the policy weaker than a licence, but stronger than a statement of concern. It creates a federal testing channel, gives agencies a 60-day deadline to build the supporting machinery and tells the biggest model developers that Washington expects earlier visibility into high-end releases.

Participation will decide whether the order changes industry behaviour. Major labs could give the government earlier insight into capabilities that are difficult to judge from model cards, launch blogs or benchmark tables. If they hold back, the order will operate mainly as a signal that the White House wants oversight without a formal permission slip.

For now, the administration has not dropped model scrutiny. It has narrowed the idea, made it voluntary and tied it to cyber risk, while keeping pre-release review in US AI policy.