US Government Orders Anthropic to Restrict Access to New AI Models

US Government Orders Anthropic to Restrict Access to New AI Models The shutdown of Anthropic’s newest AI models, Fable 5 and Mythos 5, has rapidly escalated from a technical concern into a global fight over who controls frontier AI and who gets to use it.

In early June, Anthropic launched Fable 5 as a public-facing cybersecurity model, built on the more powerful and tightly controlled Mythos 5, which had been limited to about 150 vetted firms and government agencies under “Project Glasswing.” Days later, the Trump administration issued an unprecedented export-control order: Anthropic was told to block foreign nationals from both models or take them offline entirely. The directive arrived on June 12 with a reported 90‑minute ultimatum, forcing the company to shut down the systems for all users, including its own foreign staff.

Officials cited unspecified “national security concerns,” reportedly triggered by Amazon researchers who told the White House they had bypassed Fable 5’s guardrails and by fears over Mythos’s automated hacking capabilities. The move marked the first time a government had compelled withdrawal of a publicly released frontier model and was cast by some analysts as a warning shot in a new era of technological geopolitics.

As Anthropic executives rushed to Washington to negotiate, they argued that similar vulnerabilities exist in rival systems and that export controls on cyber tools have historically failed, from encryption to spyware. Financial Times and Ars Technica reporting suggested the company’s own aggressive rhetoric about AI risk may have helped precipitate the crackdown, with critics saying Anthropic had effectively “talked itself into an AI export ban.”

Outside Washington, fallout was immediate. Developers and small businesses abroad saw projects stall overnight when access to Fable vanished mid-task, underscoring the danger of relying on a single US-hosted model. Cybersecurity experts signed an open letter warning that pulling advanced defensive tools from legitimate users could backfire, while some in the AI community took the ban as fuel for a shift toward locally run models: “Fable is banned. Long live local AI,” one entrepreneur’s post declared.

Strategists and regulators read the decision in divergent ways. The Economist argued that by deciding “who may use the world’s most important technology,” the US government had revealed itself as gatekeeper to both frontier models and the compute that powers them, a vast extension of American power. European commentators, by contrast, treated the order as a “warning shot” that Europe must respond to with its own industrial and regulatory strategy.

Across the industry, the clash has exposed AI’s missing referee. Axios described the Fable fight as turning abstract debates over oversight into a real-time test of government power. TechCrunch and others questioned whether the move was a narrowly targeted safety measure, a politically tinged dispute between Trumpworld factions and Anthropic, or even a kind of darkly beneficial marketing for the startup as the “bad boy” of AI.

Amid the arguments, one technical critique has gained traction: that post‑hoc API “guardrails” are too brittle to contain cutting‑edge systems. “They don’t make dangerous capabilities disappear. They just hide them behind a brittle interface that can be easily jailbroken,” wrote Hugging Face CEO Clément Delangue, calling instead for limits on what frontier models are trained to do in the first place. Whether Washington’s confrontation with Anthropic leads to that kind of deeper safety regime—or simply drives powerful models into new jurisdictions and onto users’ local hardware—now hangs over the next phase of the AI race.

Continue reading https://foxvector.com/stories/019ef059-aaf4-2ed6-73e7-3f47b476bf87