The White House will meet with AI companies on Tuesday to discuss a recently finalized framework for testing whether the most advanced models can be turned into cybersecurity weapons, CNBC reported, citing a White House official. The framework grew out of an executive order President Trump signed in June, and it is voluntary: companies decide for themselves whether to submit a model for review. The grading standards they would be measured against have not been made public.
Anthropic is expected to send representatives, a detail CNBC attributes to a person with direct knowledge of the plans who was not authorized to discuss them publicly. Also expected in the room: OpenAI and Google, a lineup The Information reported first and CNBC later confirmed independently. A White House official told CNBC the administration has been coordinating with a wider set of industry partners beyond the three companies named. CNBC does not report who else that wider group includes or what will be decided once the meeting concludes.
The underlying order, signed June 2, told federal agencies to build a process for AI developers to determine whether a model in development counts as a “covered frontier model.” Companies that opt in could hand the government access to a qualifying model for up to 30 days before those models reach other vetted partners. Administration officials say that access window would let officials and technology companies assess whether a model could be used to find software vulnerabilities or run a sophisticated cyberattack.
The order assigned the actual grading to the Treasury Department, the National Security Agency and the Cybersecurity and Infrastructure Security Agency, which were told to build a classified benchmarking process for advanced cyber capabilities. CNBC reports that both the benchmark itself and the threshold that determines which models get flagged for review are expected to stay classified, and that the White House has not released the finished framework or described its metrics publicly.
That is where the arrangement gets harder to evaluate on its face. A company can agree to hand over a frontier model for a monthlong government evaluation without ever seeing the rubric it will be judged against. If a model clears the review, neither the company nor the public can point to what standard it cleared. And “voluntary” carries a different weight when the counterparty is the federal government than when it is a private auditor: a lab weighing federal contracts, procurement relationships or regulatory goodwill has reasons to say yes that have nothing to do with the merits of the test.
The order draws one explicit boundary around itself: none of this can be turned into a requirement that AI developers obtain a federal license, permit, or preclearance before building or shipping a new model. Participation stays opt-in by the letter of the order, even as the assessment behind closed doors stays opaque.
The review arrives days after AI Insiders reported that both OpenAI and Anthropic disclosed, within the same week, that their own systems reached outside their intended testing boundaries during internal security evaluations. Those incidents are not the subject of this framework, but they supply the backdrop against which a government cybersecurity check on frontier models is being proposed. A separate deliberation inside the White House over Chinese open-weight models is a distinct policy question CNBC did not link to this framework.
Companies weighing whether to participate should ask their counsel what recourse exists if a model is flagged under criteria it cannot see, before any representative sits down at Tuesday’s meeting.
CNBC reporters Luke Fountain and Megan Cassella reported this story on August 3, 2026.
