The Trump administration has reversed its export controls on Anthropic’s Claude Fable 5 and Mythos 5, the company’s most advanced AI models, barely three weeks after ordering them locked down.
The speed of the about-face reveals just how chaotic the government’s approach to AI regulation has become, and how much power the companies themselves hold over the process.
What Happened, and Why It Matters
Anthropic announced on July 1 that the Department of Commerce had formally lifted export restrictions, and the company began restoring global access to both models immediately. Al Jazeera reported that the reversal came after Anthropic agreed to strengthen internal safety measures, including a new filter that blocks the specific jailbreak technique that triggered the ban in more than 99% of attempts.
The original ban, imposed in mid-June, went further than any previous AI export action. It barred not just foreign customers but foreign employees of the company itself from accessing the models, a restriction so broad that it disrupted Anthropic’s own research operations and drew immediate pushback from the tech industry, civil liberties groups, and allied governments.
The Jailbreak That Started Everything
The catalyst was a finding by Amazon researchers who discovered a technique to bypass Fable 5’s safety guardrails, prompting the model to identify software vulnerabilities. The government treated this as a national security risk and moved to restrict access across the board, according to Forbes, rather than working with Anthropic on a targeted fix first.
That sequence of events is the real story. The administration’s instinct was to impose a blanket ban first and negotiate later, treating a fixable software vulnerability like a weapons proliferation crisis. Anthropic had already been navigating the fallout for more than a week before the government came back to the table, and the eventual resolution looks almost exactly like what a cooperative approach would have produced on day one.
What the Reversal Tells Us About AI Policy
The three-week saga exposes a fundamental gap in how Washington handles AI governance. There is no standing framework for evaluating AI safety incidents, no clear escalation protocol, and no institutional memory from case to case. Each new model release creates a fresh opportunity for ad hoc decision-making, and the results are predictably inconsistent.
Anthropic came out of this episode with a stronger safety posture and a demonstrated willingness to cooperate with regulators. But the cost was real: three weeks of disrupted service, reputational damage in international markets, and a chilling effect on researchers who now know their tools can be yanked without warning based on a single vulnerability report.
The Competitive Landscape Shifts Again
The timing of the reversal is not lost on Anthropic’s competitors. OpenAI and Google DeepMind have not faced comparable export restrictions on their latest models, despite similar (and in some cases more severe) documented vulnerabilities. The asymmetry raises uncomfortable questions about whether the government’s enforcement posture is shaped by genuine security assessments or by the relative political connections of different AI companies.
For now, Fable 5 and Mythos 5 are back online globally, and the immediate crisis is over. But the precedent is set: the US government has demonstrated it can shut down access to frontier AI models on short notice, and that it will reverse course just as quickly when the economic and diplomatic costs become apparent. Neither outcome is reassuring.