Photo-Illustration: Intelligencer; Photo: Getty Images

Late last year, relations between the United States government and Anthropic began to break down. Anthropic was one of a handful of companies that had recently won contracts with the Department of Defense; unlike its peers, however, the firm sought guarantees about how its tools would be used. Domestic mass surveillance was off the table, it said, as were fully autonomous weapons. The DoD balked, Anthropic insisted, and at some point, everyone involved in these negotiations made it clear that they found the people across the table to be stupid and/or annoying. By February, the situation came to a head, publicly pitting logorrheic Anthropic CEO Dario Amodei against peacocking Defense Secretary Pete Hegseth — a match made in negotiation hell — and the government began the process of declaring Anthropic a supply-chain risk, an unprecedented and punitive move typically reserved for companies tied to hostile foreign governments. So far, so bad for the nascent project of American AI regulation.

The supply-chain designation went to court, where it remains unresolved; in the meantime, Anthropic announced a new family of models, which the company called a “step change” in cybersecurity capabilities. The company reached out to a range of peer firms and government officials to offer early access to cyber “defenders.” In the process, relations with the administration, at least some members of which appeared to be taking AI cybersecurity risks seriously, seemed to thaw. In June, Anthropic released Fable, a version of its Mythos-grade model with cybersecurity limitations built in. It was quickly metabolized like other recent AI releases: by reviewers identifying its strengths and weaknesses; by AI watchers adjusting their predictions about where the technology could be going; and by numerous customers, including anyone paying for its $20-per-month usage plan, who shared their experiences online.

Within a few days of Fable’s release, though, Anthropic withdrew access to Mythos and Fable. The company told its customers it was as surprised as they were:

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.

Behind the scenes, according to the most recent reporting, Amazon CEO Andy Jassy called Treasury Secretary Scott Bessent after engineers at his companies found a “jailbreak,” or a way to circumvent protections put in place by Anthropic to prevent misuse. Anthropic suggested that the jailbreak was neither significant nor unique to Claude and that the same results could be produced with ChatGPT; the administration, now leaking to the press, accused the company of failing to “honor” a recent, and vague, executive order about AI, which outlined a voluntary review program for new AI models and accused the company of being a “bad actor.” Hegseth couldn’t help himself:

Three months ago, @DeptofWar kicked @AnthropicAI out of our building—forever.

Every passing day proves why that was the right move. 🇺🇸

— Pete Hegseth (@PeteHegseth) June 13, 2026

This is not true, for what it’s worth. And a lot of other specifics about this situation are sort of weird at first glance. Why would Amazon, Anthropic’s largest corporate investor and a major compute partner, want to trigger a fight like this? (Could it be related to the company’s plans to become less dependent on third-party providers for AI?) Why would an administration that’s been reluctant to embrace AI “safety” as a concern suddenly wield it against the most safety-obsessed AI company?

According to Axios, the situation can be explained by some of the same dynamics present since late last year, quoting a source who says that Anthropic “has not done a great job at trying to speak to the administration and appreciate the ideological differences” and that “it’s like they just speak in different languages.” No doubt this matters: Even without the Trump administration’s vindictive, personalist tendencies, and setting aside the genuine and widely felt uncertainty about what AI regulation should look like in 2026, Anthropic is probably a particularly strange counterparty. For starters, it is led by guy who has extremely specific ideas about how the looming threats of his company’s products should be handled, not just by his firm but by regulators, accompanied by broadly liberal but deeply idiosyncratic politics informed by years of in-the-weeds discussions with AI researchers, rationalists, and effective altruists about AI alignment and x-risk. It’s easy to imagine how meetings between Trump officials and superintelligence-pilled Anthropic employees could go wrong. Honestly, it’s hard to imagine how they’d go right.

The bigger issue, though, is that they shouldn’t really need to. The administration’s approach to the AI industry, since the beginning of Donald Trump’s second term, has been fairly hands off, consistent with its generally skeptical stance on regulations of any kind, and led by a group of people with financial ties to the industry and a shared belief that American AI firms are in an arms race with China that must be won at all costs. This left a lot of space for AI firms to float the need for broad AI regulation themselves, either in earnest, as a way to garner goodwill as public opinion begins to turn against them, or as a way to get in front of backlash-driven regulations that they might find cumbersome. They’ve floated everything from new taxes and UBI programs to, as Amodei recently suggested, an FAA-like oversight structure, involving “technical testing and auditing” and in which models could be “blocked or reversed as a threat to public safety if they do not meet high standards of safety.”

In 2026, with high economic stakes, they — and the broader public — are getting something superficially similar to that but worse and enforced with little transparency or process, no predictability, and more than a bit of open spite: an informal, incoherent, and opaque oversight regime responsive only to the president, his hawkish foreign-policy team, and, maybe, occasionally, a tip called in by a business leader who may or may not have his own interests in mind. If you’re Anthropic, you get scrutiny for offering your product to a Korean firm with vaguely alleged ties to China. If you’re Elon Musk, you get a different sort of meddling: On Monday, the Justice Department intervened on behalf of xAI, which recently signed a deal to provide compute to Anthropic, claiming that a lawsuit alleging that the company’s gas-turbine-powered data centers in Mississippi violate the Clean Air Act threatens “American national, economic, and energy security by seeking to shut off the power supply for artificial-intelligence innovation that supports the Department of War’s military operations.” (The administration’s numerous previous attempts to govern by executive order haven’t worked particularly well, but it’s possible that the combination of the AI industry’s relative newness and its self-implication in national security make it a perfect target for the intervention–by–Truth Social–post governing style.)

The Trump administration’s attitude toward the AI industry up until now has been, for the most part, to let it rip. Now, with IPOs on the horizon, and talk of trillion-dollar valuations, the administration is suddenly taking a keener interest in what these companies are doing, not to make sure they’re operating responsibly, or in the broad national interest, or to ensure their products don’t cause too much economic pain elsewhere in the economy, but rather to let the industry know, bluntly, who is really in charge.

Sign Up for John Herrman column alerts

Get an email alert as soon as a new article publishes.

Vox Media, LLC Terms and Privacy Notice