The State Stopped Asking Whether AI Was Safe

The Pentagon spent the spring telling the world that Anthropic was a national security risk. Defense Secretary Pete Hegseth demanded the lab strip the guardrails that keep its Claude models out of mass surveillance and fully autonomous weapons, and when CEO Dario Amodei refused, the administration applied a "supply-chain risk" designation — a label normally reserved for foreign adversaries — and ordered every federal agency to stop using the company's technology ^[1]. Then the military kept using Claude anyway, for intelligence and targeting during operations in Iran and the capture of Nicolás Maduro in Venezuela ^[1]. The state did not think the tool was dangerous. It thought the lab was insubordinate. That distinction is the hinge on which a year of AI policy has turned. In May, Trump canceled a planned executive order that would have required labs to hand frontier models to the Treasury and the NSA up to 90 days before public release for cybersecurity vetting. Elon Musk, Mark Zuckerberg, and David Sacks had lobbied against it; Trump called it a blocker on growth and said he would not do anything to jeopardize the American lead over China ^[2]. Weeks later he signed a National Security Presidential Memorandum directing the military and intelligence agencies to accelerate AI adoption for warfighting, with Hegseth told to update autonomous weapons policy within 90 days ^[3]. One directive was replaced with its opposite: the rule that would have made the government check models before release was gone, and the rule that pushes models into the military was in. The Anthropic fight made the new logic explicit. Hegseth dismissed the lab's restrictions as nothing more than ideology.

WE will decide the fate of our Country — NOT some out-of-control, Radical Left AI company run by people who have no idea what the real World is all about. — Donald Trump

When the company sued, the D.C. Circuit upheld the blacklisting, ruling that the government's need to secure vital technology during an active conflict outweighed the lab's financial harm ^[4]. A court had now endorsed the idea that a lab's refusal to build weapons-grade capability is a security problem the state may punish. The labs, meanwhile, have begun rebuilding the containment concept — repurposed. Anthropic's Project Glasswing and OpenAI's Trusted Access for Cyber both gate access to models that can find software vulnerabilities on their own, handing them to a small circle of vetted users: Microsoft, Google, Apple, and JPMorgan on one side; thousands of cleared professionals and Five Eyes intelligence partners on the other ^[5]. "Safe" has been redefined from "safe to release" to "safe to keep secret." What remains of the old framework is thin. This week, the administration's answer to a year of existential-risk warnings from the labs' own researchers arrived: a review process that is voluntary and partly classified ^[6]. Mandatory safety vetting is gone; mandatory military integration is not. The resistance is real, and it has not changed the direction of travel. A lower-court judge, Rita Lin, found the blacklisting was likely retaliation for protected speech and put the point plainly.

If the concern is the integrity of the operational chain of command, the Department of War could just stop using Claude. — Rita F. Lin

Anthropic sued, OpenAI amended its Pentagon deal after a user exodus, and a researcher resigned over lethal autonomy ^[7]. In Congress, Bernie Sanders and Greg Casar have proposed banning superintelligence outright, with penalties modeled on weapons-of-mass-destruction law ^[8], and more than 1,500 researchers and public figures have signed a petition demanding a global ban ^[9]. None of it has stopped the substitution. And it is not an American habit. China is building the same model from its side: export controls on AI, travel restrictions on talent, a forced unwinding of the Meta-Manus deal, and government sign-off for foreign investment in AI firms ^[10]. Both superpowers have stopped treating frontier AI as a product to be regulated and started treating it as a strategic asset to be controlled. The state has not abandoned control of frontier AI. It has changed what "control" means. The question is no longer whether a model is safe to release. It is who gets the dangerous version.