Inside the whirlwind 24 hours that led the White House to slap export controls on Anthropic

{{IMAGE:1}}

Anthropic has spent years positioning itself as the AI company that wants regulation. CEO Dario Amodei has compared AI risks to nuclear weapons. The company has pushed for government oversight, submitted to voluntary reviews, and advocated for the kind of safety guardrails that the industry's more libertarian voices resist. So when the Trump administration moved to impose export controls on Anthropic's Fable 5 model last week, the company found itself in an unfamiliar position: arguing against the very regulatory framework it had helped popularize.

The sequence of events, reconstructed from administration officials, Anthropic representatives, and people familiar with the discussions, reveals how quickly concerns about a single jailbreak vulnerability escalated into an unprecedented government action against a major AI developer. It also exposes the fragile equilibrium between innovation speed and safety verification that the AI industry has yet to resolve.

The Trigger: Amazon's Security Findings

Fable 5 launched publicly earlier this week, described by Anthropic as a "Mythos-class model" with safeguards designed to make it safe for general use. The model had undergone reviews by both the administration and the United Kingdom's AI Security Institute. Anthropic believed it was ready for broader deployment.

But on Thursday, two days after the public release, Amazon CEO Andy Jassy raised concerns to the White House. Amazon, an investor in Anthropic, had been responding to an administration request for feedback and had identified what it believed was a way to bypass the model's guardrails. The specific nature of the bypass wasn't immediately clear from public reporting, but the administration treated it as a serious national security matter.

Amazon's findings were subsequently validated by the National Security Agency, according to a White House official. The NSA's involvement elevated the issue from a product concern to a potential national security threat, giving the administration both the justification and the urgency to act.

The Escalation: From Meeting to Export Controls in 24 Hours

By Friday morning, the issue had reached the highest levels of the White House. Treasury Secretary Scott Bessent, AI policy advisor Michael Cairncross, Chief of Staff Susie Wiles, and other senior officials convened to discuss the model and potential responses. Bessent, traveling to Houston for a public event, joined remotely.

The administration attempted to reach Amodei but was initially told he was unavailable. A spokesperson for Anthropic disputed the administration's claim that Amodei was attending a wellness retreat, calling it "absolutely false." A person close to Anthropic said Amodei was first contacted around noon and was on the phone with senior officials within an hour and fifteen minutes. Anthropic offered other senior leaders in his absence, but the administration wanted to speak directly with the CEO.

When Amodei finally joined, he participated in three calls with a rotating cast of roughly half a dozen senior administration officials. The group included Cairncross, Bessent, Commerce Secretary Howard Lutnick, Under Secretary of Commerce for Industry and Security Jeffrey Kessler, White House Staff Secretary Will Scharf, Deputy Chief of Staff Richard Walters, and Assistant to the President for Policy Walker Barrett.

Amodei's position was that the bypass was specific and limited, not the kind of universal jailbreak that would render the model's safety systems entirely useless. He argued that the type of vulnerability identified didn't pose the same risk as a broad compromise that would allow unrestricted use. In a blog post published after the export controls were imposed, Anthropic stated that "no testers have yet been able to find a universal jailbreak" and that complete prevention of all jailbreaks wasn't currently possible for them or any other company.

The administration wasn't convinced. Bessent told Amodei directly that he was making a "bad decision," according to a senior White House official. Cairncross and Bessent pressed Anthropic to voluntarily remove the model and coordinate with the government on a remediation plan. Amodei asked for more time and information but made no commitments.

The Nuclear Option: Export Controls

Shortly after the calls concluded, the administration imposed export controls on Fable 5 and Mythos 5, citing national security authority. The order banned use by foreign nationals, which Anthropic said had the "net effect" of "abruptly disabling" the models for all customers to ensure compliance.

The choice of export controls as the enforcement mechanism was notable. Rather than a direct product ban or a demand for specific fixes, the administration reached for a tool typically reserved for dual-use technologies with military applications. The move effectively treated advanced AI models as controlled technology, similar to semiconductor manufacturing equipment or cryptographic systems.

"Export controls were a last resort after begging them for hours to work with us," a senior White House official said. "This was not something we wanted to do, but our hands were tied."

Anthropic contested this framing. A person close to the company said the White House gave Anthropic 90 minutes to take the models down, with no details on the actual threat. "There was never any begging, or asking, for them to work with us, just a declared 90 minute deadline," the person said.

The Irony: Regulation's Proponent Regulated

The episode places Anthropic in an awkward spot. The company has been among the most vocal advocates for AI regulation in the industry. It has participated in government-led safety reviews, pushed for mandatory testing requirements, and supported the kind of executive order that established the current framework for advanced AI oversight. Amodei himself has warned repeatedly about the risks of uncontrolled AI development.

But when regulation was applied in a way the company disagreed with, Anthropic pushed back. In its blog post, the company called the export controls "disproportionate" and said it believed "the government should have the ability to block unsafe deployments, as part of a statutory process that is transparent, fair, clear, and grounded in technical facts. This action does not adhere to those principles."

This tension reflects a broader debate within the AI safety community. Many researchers and executives support the concept of regulation in principle but become uncomfortable when it constrains their specific products or timelines. The gap between abstract support for oversight and concrete acceptance of its consequences has always been wide. Anthropic's experience suggests that gap may be unbridgeable under the current political dynamics.

The Political Context

The administration's decision didn't happen in a vacuum. Anthropic has had a contentious relationship with the Trump White House for months. Secretary of Defense Pete Hegseth and the Pentagon designated the company a supply chain risk in March over its refusal to allow its AI tools to be used for mass domestic surveillance and autonomous weapons. The administration has accused Anthropic of leftist political bias and fearmongering through its advocacy for stronger regulation and warnings about mass job disruption.

David Sacks, the former White House AI czar and a vocal opponent of regulation, nonetheless supported the administration's decision. "The Admin values Anthropic's technical capabilities and feels that this issue, while serious, should be easily resolved," Sacks wrote on X. He argued that the jailbreak was neither simple nor trivial and that the export controls weren't an attempt to exert broader control over the industry.

Sacks' support is telling. It suggests the administration's action wasn't purely punitive or politically motivated, even if the underlying relationship between Anthropic and the White House was already strained. The decision appears to reflect a genuine belief that the security vulnerability posed a real risk, combined with frustration that Anthropic wasn't taking it seriously enough.

The Precedent: What This Means for AI Governance

The Anthropic episode sets several precedents that will likely shape AI governance for years to come.

First, it demonstrates that export controls can be applied to AI models, not just hardware. This expands the government's toolkit for managing AI risks beyond traditional product liability or sector-specific regulation. It also raises questions about how such controls would be enforced, especially for models that can run locally or on-premises.

Second, it reveals the limits of voluntary compliance frameworks. Anthropic had participated in pre-release reviews and believed it had satisfied the government's requirements. The sudden imposition of export controls after those reviews suggests that the current process lacks clear triggers and escalation paths. Companies may now face uncertainty about whether their compliance efforts will be sufficient.

Third, it exposes the tension between speed and safety in AI deployment. Anthropic wanted to move quickly with Fable's public release. The administration wanted more time to verify the model's safety. Neither position was unreasonable, but the lack of established procedures for resolving such disputes led to an outcome that neither side wanted.

What Comes Next

The administration has signaled that it wants Anthropic to fix the safety issue and have the export controls lifted. Sacks wrote that the "ball is in Anthropic's court" and that the administration hopes the remediation happens quickly.

Anthropic faces a difficult choice. Complying with the administration's demands would validate the government's approach and establish a precedent for future interventions. Fighting the export controls could lead to a prolonged legal battle and damage the company's relationship with regulators at a critical moment.

The broader AI industry is watching closely. If Anthropic capitulates, it may encourage the administration to take similar actions against other developers. If Anthropic resists successfully, it could embolden other companies to push back against regulatory demands. Either outcome will shape how AI governance evolves in the coming months.

The deeper question is whether the current framework for AI oversight can accommodate the speed at which the technology is advancing. The Anthropic episode suggests that it cannot. The gap between deployment decisions and security verification is too wide, the escalation pathways are too abrupt, and the stakes are too high for the ad hoc approach that characterized last week's events.

For an industry that has spent years debating how to regulate AI, the Anthropic case offers a concrete example of what happens when those debates remain unresolved. The answer, it turns out, is a 24-hour escalation from security concern to export control, with little room for the nuanced analysis that both sides claimed to want.