Washington Runs Frontier AI by Guest List, Not Law

Washington is governing America's frontier AI through improvised kill switches and customer-by-customer guest lists — while the executive order that was supposed to build a real framework won't deliver one until August at the earliest.

What's new: In two weeks, the Trump administration intervened in both frontier labs' flagship releases — with different agencies, different legal tools, and no public rulebook.

• On June 12, Commerce Secretary Howard Lutnick ordered Anthropic to disable Fable 5 and Mythos 5 worldwide under export-control authorities, per CNBC. Anthropic said it got 90 minutes to comply and received no specific details of the national security concern, per CNN.
• On June 26, the White House's Office of the National Cyber Director and Office of Science and Technology Policy asked OpenAI to limit GPT-5.6 Sol, Terra, and Luna to roughly 20 government-vetted partners, per Axios.
• That same Friday, Lutnick partially restored Mythos 5 — but only for roughly 100 organizations on a federal allow list. Fable 5 stays dark.

Why it matters: The interventions look coordinated. The governance is not.

• President Trump signed Executive Order 14409 on June 2, calling for a voluntary 30-day pre-release framework and explicitly disclaiming "mandatory governmental licensing, preclearance, or permitting."
• Agencies have 60 days — until early August — to design the classified benchmarking process that defines a "covered frontier model."
• In the gap, frontier release runs on improvised clearance: export law for Anthropic, White House vetting for OpenAI, and no published standard for either.

What the labs said: Both complied. Neither endorsed the arrangement.

• OpenAI wrote in its launch post: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."
• CEO Sam Altman told staff Washington would approve access "customer by customer" during preview, per Axios. His memo: "We've made clear to the U.S. government that this is not our preferred long term model."
• Former White House AI adviser Dean Ball — soon joining OpenAI — told TechCrunch the June 2 order has created a "de facto involuntary licensing regime" when paired with undefined safety standards.

What the critics said: Even safety advocates think this is broken.

• Brad Carson, head of the bipartisan pro-AI-safety super PAC Public First, told CNN: "Right now, you have an ad hoc, personalized, opaque, possibly lawless approach."
• Jessica Tillipman, George Washington University associate dean for government procurement law, wrote: "What is striking is the absence of any meaningful process." She warned an administration governing this way "will not avoid the heavy regulation it fears" — per CNN and her essay.
• Just Security notes Commerce has never before applied export controls to a live commercial AI model — and has not published the Anthropic directive, the legal basis, or remediation steps.

The structural mess: Three problems stack on top of each other.

• Agency roulette: Anthropic's shutdown came from Commerce. OpenAI's staggered rollout came from the White House. The June 2 EO assigns NSA to classify which models count as frontier. Nobody has published who is in charge when they disagree.
• Tool mismatch: Export law was built for chips crossing borders — not API access where foreign nationals inside the U.S. trigger a global kill switch, per TechPolicy.Press and Just Security.
• Evidence vacuum: Anthropic said the flagged jailbreak involved minor vulnerabilities common to other models. Former Facebook CISO Alex Stamos wrote on X: "There were some valid findings but no unique capabilities that justify a reaction close to this." Trump AI adviser David Sacks countered: "It's difficult to fathom how they could claim a jailbreak allowing operability of a cyber weapon could be defined as not 'serious.'" — both cited by CNN.

What to watch:

• August 1: The EO's deadline for NSA's classified cyber-capability benchmark and Commerce's voluntary framework — or another month of guest-list governance.
• Rep. Nathaniel Moran's AI Incident Reporting Act: A seven-day notification clock for dangerous model behavior — the scalpel Congress proposed two weeks after Commerce used the sledgehammer.
• China's open-weight releases: GLM-5.2, Kimi K2.7, and Qwen 3.7 shipped as downloadable weights while U.S. frontier models clear Washington one customer at a time.

Convina's view: Deregulation was the promise. Improvisation is the product. Washington is not governing frontier AI — it is rationing it through private letters, unpublished export orders, and annex lists that can change at any time. That may buy a news cycle of safety theater. It will not buy the predictability that $965 billion IPOs, allied governments, and enterprise buyers all need before they sign the next contract.