White House Meets AI Firms, Keeps Model Rules Secret

A Closed-Door Meeting on Frontier AI Safety

The White House convened a high-profile gathering with major AI developers—including Meta, Nvidia, Microsoft, OpenAI, and Anthropic—along with a range of smaller companies to review a new framework for evaluating AI models, but declined to make the document public 1. The meeting produced no public commitments, and officials have not clarified whether they intend to again exercise the authority that previously forced two Anthropic models offline in June 6.

The framework reportedly traces back to a June 2 executive order signed by President Trump, which directed federal agencies to establish a voluntary process for AI developers to self-determine whether their models qualify as "covered frontier models" 4. Despite completing the framework, the administration has kept its specific rules and criteria shielded from public view, leaving developers, researchers, and watchdogs without clarity on how frontier-model risk will be assessed going forward 4.

Why the Secrecy Raises Questions

The timing of the closed-door session is notable: it follows a string of security incidents involving OpenAI and Anthropic models that have unsettled public confidence in frontier AI safety 1. Given that backdrop, the decision to withhold the evaluation framework—rather than publish it for scrutiny—has drawn attention from observers who note the apparent tension between promoting voluntary industry cooperation and maintaining opacity around the rules governing it 16. The earlier episode in which the government forced Anthropic models offline underscores that federal regulators already possess real enforcement leverage, even under a nominally voluntary framework, making the lack of transparency more consequential 6.

The Broader Competitive Backdrop

The policy discussion arrives amid a fast-moving wave of model releases that illustrate just how competitive the frontier AI landscape has become. Microsoft has introduced a new cybersecurity-focused model, MAI-Cyber-1-Flash, which it claims outperforms Anthropic's Claude Mythos 5 on security benchmarks while costing roughly half as much to run 25. Independent tracking of new model releases has emerged as its own cottage industry, reflecting how frequently competing labs are leapfrogging one another on performance claims 2.

International competition is intensifying as well. Alibaba unveiled its Qwen3.8-Max model, explicitly positioning it as a rival to leading systems from OpenAI and Anthropic and signaling that the global race for AI supremacy now extends well beyond U.S. borders 3.

What It Means

Taken together, the coverage points to a widening gap between the pace of commercial AI development—marked by rapid, competitive model launches from Microsoft, Alibaba, and others—and the pace of regulatory transparency. While the administration frames its approach as voluntary and collaborative, the refusal to disclose the evaluation framework, combined with its demonstrated willingness to pull models offline, leaves open questions about how frontier-model oversight will actually function, and how much visibility the public or independent researchers will have into decisions that could shape which AI systems reach the market.