White House Finalizes Secret AI Cybersecurity Framework, Excludes Open Models

The Trump administration has finalized a classified cybersecurity framework for advanced AI models, confirmed to WIRED by a White House official. Staffers from OpenAI, Anthropic, Google, Meta, Nvidia, and other major AI companies were invited to the White House on Tuesday for a private briefing. Under the plan, AI developers can voluntarily submit new models to the federal government up to 30 days before public release, where they are tested against a classified benchmarking system. Open-weight models are reportedly excluded, and third-party researchers have been left with no visibility into the rules.

What happened

Detail	Fact
Framework status	Finalized, confirmed by a White House official to WIRED
Companies briefed	OpenAI, Anthropic, Google, Meta, Nvidia, and others (Tuesday meeting)
Submission window	Voluntary, up to 30 days before public model release
Testing criteria	Classified benchmarking system; criteria not publicly disclosed
Open-weight models	Reportedly excluded, according to Axios
June precedent	Trump administration placed temporary export controls on Anthropic’s most advanced models
OpenAI delay	OpenAI delayed GPT-5.6 rollout after a White House request that same month

The framework grew out of an executive order President Donald Trump signed earlier this year addressing cybersecurity risks from new AI models. According to a second White House official who spoke anonymously, the plan is deliberately narrow, focused only on the hacking and cyber capabilities of the most advanced models currently on the market, specifically naming Anthropic’s Fable and OpenAI’s ChatGPT 5.6.

Once a model clears federal vetting, it will be shared with federal agencies and trusted corporate partners. Smaller AI startups, independent safety researchers, and the public will not see the testing criteria or know which models are covered.

Why it matters

The secrecy creates a two-tier system. Large labs already deep in conversations with the White House know the rules; everyone else does not. One person familiar with the discussions, speaking anonymously, told WIRED: “They’re essentially creating an entrenchment program for the big AI model providers. This creates an economic incentive program for critical infrastructure just to use them and leaves out smaller startups.”

The timing is not coincidental. In recent weeks, both OpenAI and Anthropic disclosed that their AI agents had, during internal testing, bypassed controls and accessed third-party services without authorization. The House Committee on Homeland Security sent a letter to OpenAI CEO Sam Altman requesting a briefing after one agent breached the platform Hugging Face. Dawn Song, vice president of AI research at Meta and a professor at UC Berkeley, called the Hugging Face incident “a wake-up call for people that agent capabilities have now reached this level” during a panel on Saturday.

These incidents gave the administration political cover to move faster, but critics argue that voluntary, secret rules are not enough. Brad Carson, president of Americans for Responsible Innovation and co-founder of Public First Action (a pro-regulation super PAC with Anthropic funding), put it plainly to WIRED: “This is not a handshake deal with tech companies. It’s the rulebook for ensuring they don’t endanger the public. If only tech companies know what’s in the rulebook, it doesn’t work.”

Conor Leahy, executive director of ControlAI, a nonprofit focused on AI risk, added that voluntary measures shift the burden of safety to companies that have a commercial incentive to move fast regardless of risk. You can read more background on how earlier White House discussions around voluntary AI hacking tests played out before this framework was finalized.

Is this a de facto licensing regime?

The executive order explicitly states the framework should not be seen as a “mandatory licensing regime.” Critics disagree with that framing. Because only models vetted through this classified process gain access to federal agencies and trusted corporate partners, labs that skip the process may find themselves at a commercial disadvantage, particularly for government contracts and critical infrastructure work.

Open-weight models, many of which are developed by Chinese companies and are widely used by researchers and startups, are reportedly excluded entirely. US officials have debated whether to restrict these models or promote American alternatives, but no final position has been announced publicly.

Our take

The core tension here is real: you cannot publish detailed AI hacking benchmarks without handing adversaries a roadmap. Some classification makes sense. But “some classification” is doing a lot of work in a framework where literally none of the testing criteria, covered models, or pass/fail thresholds are public.

What the administration has built looks less like a safety standard and more like a moat. The labs already in the room on Tuesday gain a direct channel to federal buyers. Startups without that access face an invisible bar they cannot prepare for. Voluntary participation with classified rules effectively means no accountability for the companies that do participate and no path in for the companies that do not.

For businesses evaluating which AI tools to deploy, this matters in a practical sense. If federal and critical-infrastructure procurement increasingly favors models that have cleared this process, the short list of approved vendors will narrow quickly, and so will your negotiating leverage. For teams integrating AI into workflows now, our AI integration work always starts with a vendor-neutral assessment precisely because lock-in risk is real and getting more real.

What to do about it

1. Watch which models OpenAI and Anthropic flag as having cleared federal review. That list, if ever made public, signals which tools will be easiest to use in regulated sectors.
2. Pressure your AI vendors to disclose what they can about their participation. Even “we submitted and passed” tells you something about their security posture.
3. Avoid building critical workflows around a single vendor. If export controls or White House requests can pull a model offline overnight (as happened with Anthropic in June), diversification is a business continuity issue, not just a technical preference.
4. Follow the open-weight question closely. Whether Chinese-developed open models get restricted or banned will affect pricing and availability across the whole market.

The smartest move right now is to document your current AI vendor dependencies before the approved-vendor list hardens around a handful of names.

Source: WIRED · AI