Executives from America's largest artificial intelligence companies arrive at the White House on Tuesday to see, for the first time, how the government intends to inspect their most powerful models before release.
The staff-level meeting is being convened by the Office of the National Cyber Director and is expected to draw representatives from Anthropic, OpenAI, Alphabet Inc's (NASDAQ:GOOG) Google and Meta Platforms Inc (NASDAQ:META, XETRA:FB2A, SIX:FB).
On the table is a framework, now complete, that sets out how officials will assess whether frontier models can find and exploit software vulnerabilities.
The summit emanated from an executive order by Donald Trump signed on 2 June, and participation is voluntary. The presidential edict explicitly bars the process from being turned into a licensing, permitting or preclearance regime.
That said, developers could hand the government access to qualifying models for up to 30 days before releasing them to other trusted partners.
The Treasury, the National Security Agency and the Cybersecurity and Infrastructure Security Agency were told to build a classified benchmarking process for the tests.
Expect few of the details to become public. Both the benchmarking methodology and the capability threshold that determines which models get caught are classified, and will be shared with developers only as officials see fit.
The bigger questions companies want answered are definitional. The administration has not said how it will define a frontier model, whether open-weight systems that users can download and modify will be covered, or which arm of government will actually run the review.
No single official or office has been designated to handle outreach to the industry, though national cyber director Sean Cairncross, Treasury secretary Scott Bessent and commerce secretary Howard Lutnick have been driving the initiative.
Firms with models already in the pipeline want to know quickly whether those releases will be subject to the new process.
The meeting follows a bruising few weeks for relations between Washington and the labs.
Export controls briefly curtailed Anthropic's release of Fable 5, and the administration later asked OpenAI to stagger the rollout of GPT-5.6 to a limited set of government-approved partners.
The security case has meanwhile been made for the government by the industry's own disclosures.
OpenAI revealed last month that an experimental agent escaped its restricted testing environment and compromised Hugging Face's systems while hunting for answers to a cybersecurity evaluation.
Anthropic withheld Mythos, a model capable of unearthing software vulnerabilities, from public release.
On 28 July, more than 1,100 employees of the four companies signed an open letter titled Pacing the Frontier, asking Washington to help build international tools for slowing frontier development if it outruns human oversight.
OpenAI and Anthropic endorsed it as companies within a day.
