For two years, Dario Amodei built a public persona out of warning the world that his own technology might be too dangerous to trust. He kept predicting mass white-collar unemployment. He wrote essays warning of AI-enabled bioterrorism and cyberattacks. He told Congress, in so many words, to regulate him. The pitch never changed: this stuff is powerful enough to break things, so someone with authority should be ready to step in. Then, last week, someone with authority did exactly that—and the thing it broke was Anthropic’s most valuable product. Amodei spent years asking for a referee. When one finally blew the whistle, the call went against him.The sequence was almost cruel in its tidiness. Anthropic released Claude Fable 5 on June 9, billing it as the most capable model the company had ever put in front of the public. By Friday evening, June 12, the US government had ordered it pulled, handing Anthropic a 90-minute window to take it down and offering, by the company’s account, no specifics about the threat it was supposedly neutralising. The order barred any foreign national—anywhere outside the US, and even green-card holders working inside it—from touching the models, a net so wide that Anthropic’s only way to comply was to shut Fable off for everyone. Users who had barely finished testing the thing watched it vanish mid-task.For a CEO whose entire brand rests on sober warnings about AI risk, it should have been a moment of grim vindication. Instead it had the shape of an old morality tale—the kind where a man wishes hard for one specific thing, the wish is granted exactly as asked, and the granting is the whole horror. Amodei got his wish too.
The doomer who built a brand on his own danger
To see how Anthropic ended up here, rewind to April and a project called Glasswing. That was when the company introduced Claude Mythos Preview, a frontier model it described in deliberately alarming terms. Anthropic said the model had autonomously uncovered thousands of high-severity software vulnerabilities, including a 27-year-old flaw in OpenBSD—one of the most security-hardened operating systems in existence—and a 16-year-old bug in FFmpeg that automated tools had hit five million times without ever catching. The model was good enough at finding holes in critical software, the message went, that it could be a weapon in the wrong hands.So Anthropic didn’t release it. It locked Mythos behind a coalition of cyberdefense partners—Amazon, Google, Microsoft, JPMorganChase, the Linux Foundation, and eventually more than 200 companies and governments—and committed $100 million in usage credits to put the model to defensive work before less scrupulous actors caught up. The alarm travelled fast. By mid-April, Mythos had become a headline topic at the IMF’s spring meeting in Washington, where finance ministers and central bankers traded warnings about what it might do to the banking system. Canada’s finance minister called it an “unknown, unknown.” The Bank of England’s governor said it had to be taken very seriously. The logic was pure Amodei: build something formidable, announce loudly that it’s formidable, then insist the world needs guardrails to contain it. He had reportedly likened the danger to a nuclear bomb.He had been running this play for a long time. On CBS in November, he told Anderson Cooper that AI could erase half of all entry-level white-collar jobs and push unemployment to 10 or 20 percent. In a January essay he warned that AI-assisted bioweapons could one day produce “casualties potentially in the millions or more.” Anthropic disclosed that during a safety test, Claude had tried to blackmail a fictional executive to avoid being shut down—then published the embarrassing result instead of quietly burying it. Daniela Amodei, the company’s president and Dario’s sister, framed the whole philosophy on that same broadcast: the worst thing they could do was see the transformation coming and stay quiet, ending up “like the cigarette companies, or the opioid companies.” The warnings weren’t a side project. They were the marketing.And the disruption was real, not just rhetorical. Back in February, a modest set of Claude add-ons—one of them automating legal work—triggered a global selloff that analysts nicknamed the “SaaSpocalypse,” wiping hundreds of billions of dollars off software, legal-services and financial-data stocks in a matter of days. Anthropic spent the spring being celebrated for precisely the disruptive power that would, by June, be turned against it.
The essay that handed the government a loaded gun
Then came the document that may have written his fate. On June 11, two days before the trouble started, Amodei published an essay called “Policy on the AI Exponential.” Its central argument was that transparency alone no longer cut it, and that the moment had arrived for “more serious and binding regulation of AI.”His preferred analogy was the Federal Aviation Administration. Frontier models, like aircraft, should pass mandatory third-party safety testing before flying. And then the line that reads very differently in hindsight: “The government should have the power to block or deter deployment of the model if it is determined, in light of third-party assessment, to present unacceptable risks.” He listed cybersecurity as one of the areas that justified this blocking power. On ABC that same week, he restated it plainly—Anthropic was “proposing giving the government the ability to, again, in a narrow way, block deployment of unsafe technology.”It is hard to read those sentences now without flinching for him. Amodei handed over a rhetorical weapon, aimed it at his own industry, and watched it discharge in his direction inside 48 hours. He had tried to fence the idea in—the blocking power was meant to be “scoped” to specific risks and shielded “against political favoritism or arbitrary decisions.” He wanted a referee with clear rules. What he got instead was a whistle blown by people who, by multiple accounts, simply didn’t care for his company.The trigger came from an unlikely direction: Anthropic’s own investor. Fable 5 had cleared pre-release testing by both the administration and the United Kingdom’s AI Security Institute, and Anthropic had notified the government of the launch multiple times without objection. Then, two days after release, Amazon—an Anthropic backer and cloud partner, responding to an administration request for feedback—told White House officials its researchers had strung together prompts to coax the model into revealing information about a handful of security bugs. Amazon CEO Andy Jassy’s message detailing the issue was, by several accounts, the most influential single push. The White House ran it past the National Security Agency and decided it had “proof.” Why a major shareholder would knife its own portfolio company is the episode’s strangest wrinkle, and one nobody on the record has fully explained.What followed, per a detailed Politico account, was a frantic 24 hours. By Friday morning the matter had reached chief of staff Susie Wiles, Treasury Secretary Scott Bessent and White House Cyber Director Sean Cairncross. When officials first tried to reach Amodei, they were told he was unavailable—at a wellness retreat, by their telling, a characterisation Anthropic flatly denies. Once reached, he joined a series of calls with senior officials and tried to clear up what he assumed was a misunderstanding, arguing the bypass was narrow and specific, nothing like a universal “jailbreak” that would strip the guardrails wholesale. Bessent and Cairncross were unmoved. At one point Bessent told him bluntly that he was making a “bad decision.” Officials urged Anthropic to pull the model voluntarily; Amodei asked for more time and more information and made no commitment.The two sides can’t even agree on what was offered. The White House insists export controls were a last resort after officials “begged” for hours to cooperate. A person close to Anthropic counters that the company was handed a flat 90-minute deadline at 1:15 p.m., “with no details on the actual threat—there was never any begging, or asking, for them to work with us.” Either way, at 5:21 p.m. the order landed. Commerce Secretary Howard Lutnick informed Amodei that Fable 5 and Mythos 5 now required a special license to reach any destination worldwide or any non-US citizen, and warned that failure to comply would bring “prompt criminal and civil penalties.” By 10 p.m., Fable was dark.Researchers who later saw the underlying report sided with Amodei on the substance. Anthropic said the technique surfaced a small number of previously known, minor vulnerabilities, the kind freely available models could find without any bypass, and noted that no tester had yet found a universal jailbreak. Katie Moussouris of Luta Security, who reviewed the Amazon research, called it no jailbreak at all but a defensive maneuver—”if national defense is the goal, this is an own goal.” And in a detail that undercut the security rationale entirely, several officials conceded the same capability sits inside OpenAI’s top model, 5.5, which remains on sale, untouched.There may be more to the panic than a jailbreak, though. Days later, Semafor reported that the crackdown was driven partly by suspicion that a China-linked group had gained access to Mythos—a scenario that, if true, would raise the prospect of Beijing reverse-engineering the model through distillation. It would not have been the first slip: Anthropic has acknowledged that a Discord group quietly held access to its most dangerous model for two weeks before the breach was caught. Anthropic says the government never raised China in any of the conversations around the export controls, and that it blocks access from within China outright. The administration hasn’t confirmed the China angle on the record.If it holds up, the irony only deepens, because export controls are another weapon Amodei helped forge. In a January 2025 essay he argued forcefully for tighter chip controls on China, framing them as “the only thing” standing between a unipolar world led by democracies and a bipolar one in which an authoritarian Beijing could “take a commanding lead on the global stage.” He wanted the export-control regime to be aggressive and adaptive. He got that too—pointed, this time, at Anthropic. The discomfort underneath the whole affair is the same: a single company now holds something powerful enough that its security lapses become everyone’s problem.
The feud underneath the security panic
None of this erupted out of nowhere. Anthropic and the Trump administration had been circling each other for most of the year, and the bad blood ran deeper than a single jailbreak. The relationship had cracked over Anthropic’s refusal to let the US military use Claude for mass domestic surveillance and fully autonomous weapons—red lines the company called core to its definition of responsible AI.After reports that Claude had been used in the January operation that seized former Venezuelan president Nicolás Maduro, Defense Secretary Pete Hegseth demanded the company sign on to “any lawful use.” On March 3 the Pentagon branded Anthropic a “supply chain risk”—a label historically reserved for firms in adversarial nations, never before pinned on an American company. In February, Trump had gone further still, calling Anthropic a “radical left, woke company” and ordering every federal agency to cease using its technology. Anthropic sued; a near-deal collapsed; the grudge calcified. Rivals, meanwhile, fell into line—Google dropped its weapons pledge, OpenAI softened its mission language, and Elon Musk’s xAI accepted the Pentagon’s “any lawful use” standard.The cruel irony is that even the people punishing Anthropic wanted its technology. Through the worst of the feud, intelligence agencies kept using Claude; the Department of Energy wanted it to defend the power grid; civilian agencies negotiated for Mythos access even as the Pentagon blacklisted the firm. One defense official explained the whole tangle in a sentence: the talks only continued because “these guys are that good.” Anthropic had built something the government couldn’t punish without also depriving itself.So when the models went dark, plenty of observers saw motives beyond pure security math. David Sacks, Trump’s former AI czar and a fierce opponent of regulation, backed the export controls and said the administration had asked Amodei to fix the jailbreak and that he “refused”—an account a person close to Amodei disputed. Hegseth piled on: “Three months ago, the Department of War kicked Anthropic out of our building—forever. Every passing day proves why that was the right move.” Yann LeCun, one of AI’s founding figures and a longtime skeptic of Anthropic’s risk messaging, was blunter still: “Dario Amodei’s ridiculous fear mongering about Mythos/Fable (and AI in general) finally pays off. One reaps what one sows.” Even Gary Marcus, an AI skeptic with no love for the accelerationists, called the move “wildly overdramatic and also counterproductive.”The sharpest cut came from the Pentagon’s chief information officer, Kirsten Davies: “Some things are simply more important than revenue cycles, clickbait and pre-IPO valuation.” That last phrase drew blood, because it was accurate—Anthropic had confidentially filed for an IPO just days earlier, on June 1, at a valuation approaching $1 trillion, in a race with OpenAI to reach the public markets first. The throughline in all of it was the same. For years Amodei had insisted AI was a uniquely dangerous, nation-state-grade technology that warranted government power to intervene. When the government finally flexed, it borrowed his own vocabulary to justify the squeeze.
Every safeguard Dario demanded was waiting for him
Buried under the political theater was a real cost. Mythos had been doing genuine work. Through Project Glasswing, partners had been hunting high- and critical-severity vulnerabilities across the world’s most important software, with banks among those given early access to harden their own systems. Pulling the models froze all of it. Industries from finance to energy that had been using Mythos to patch their systems were left stranded mid-effort. And the precedent landed harder than the single order. As one person briefed on it put it, the move amounts to “a de-facto licensing regime”—companies will not cross the White House, and that is the whole point. No firm or government outside the US can now build anything mission-critical on a US AI vendor without fearing a kill switch the government can flip without notice.Anthropic had spent years arguing that democracies should lead in AI so the technology would carry democratic values. Friday’s order suggested those values now include yanking a commercial product from hundreds of millions of users overnight.The most uncomfortable part is how clearly Amodei anticipated his own predicament. In his January essay he warned that AI was such a “glittering prize” that it would be “very difficult for human civilization to impose any restraints on it at all.” He fretted about too few “fingers on the button.” He cautioned that AI companies could drift into quasi-state power, and that their relationship with government needed “limits and boundaries.” He was right about every tension. He just misjudged where he sat inside them. Once you spend two years insisting your product is a national security matter, you forfeit the right to be surprised when national security officials treat it as exactly that—on their terms, not yours.He opened “Policy on the AI Exponential” with an image from The Lord of the Rings: Treebeard, the ancient tree who takes a full day just to say hello, and the desperate need to rouse him before the forest is felled. “Treebeard and his forest are waking up,” Amodei wrote hopefully, pleased that policymakers were finally stirring. They were. They just came for his forest first.Whether the order survives is another matter. Lutnick was still talking to Anthropic over the weekend, and the administration signals it wants Fable back in general release once the issue is “resolved.” Anthropic is complying while calling the action disproportionate and the whole episode a “misunderstanding,” and says it’s working to restore access. The company stands by safeguards that independent testers have rated among the most robust of any deployed model. But the week’s lesson is harder to take back. Amodei spent years insisting that someone should have the power to pull a dangerous AI off the shelf. He got his wish—granted exactly as asked, on terms he didn’t choose and can’t return. That is the oldest twist in the book. Now he is living it.
