Inside the U.S. Government's Unprecedented Shutdown of Anthropic's Mythos AI

In its own system card, Anthropic called the model’s capabilities a “cybersecurity reckoning,” noting it could identify “high-severity vulnerabilities” in major operating systems and web browsers at a level that made public release irresponsible . The company chose instead to gate access behind an invitation-only program called Project Glasswing, which by June 2026 had grown from roughly 40 vetted partners to about 200 organizations, including Amazon Web Services, Microsoft, JPMorgan Chase, and CrowdStrike .

Why Anthropic blocked its own general release

From the moment of its announcement, Anthropic said it had no plans to release Mythos to the public . The stated reason was straightforward: a model this good at finding zero-day exploits could be weaponized by malicious hackers or state-sponsored actors to attack critical infrastructure at scale . The company framed the restriction as “defensive acceleration” — giving defenders a head start before attackers could catch up .

This was not a hypothetical. In late April 2026, Anthropic confirmed it was investigating reports that unauthorized users had gained access to Mythos through a third-party vendor environment . The incident reinforced the company’s stance that general availability was incompatible with safety.

The brief life of Fable 5 and Mythos 5

On June 9, 2026, Anthropic changed course — partially. It released Fable 5, a “safe” public-facing model from the Mythos class with limitations on sensitive applications . The company described it as a way to let the public test Mythos-class technology while retaining guardrails that blocked misuse. At the same time, Anthropic offered the unrestricted Claude Mythos 5 to its existing vetted partners .

The launch was widely covered as a cautious step toward openness. The Guardian reported that Fable 5 carried “limitations on its application in sensitive fields,” while Fortune described it as Anthropic “making its first Mythos-tier model available to the general public” . The distinction between a safe public version and a partner-only full model appeared to be Anthropic’s strategy for managing risk.

The June 12 order and total shutdown

At 5:21 p.m. ET on June 12, 2026 — three days after Fable 5’s launch — the Trump administration issued an export control directive to Anthropic. The order required the company 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 Commerce Department sent the letter, but the government cited broad “national security authorities” rather than any specific statute . The order’s reach was extraordinary: it applied to foreign nationals physically present in the U.S., including people who had helped build the models . In U.S. export control law, this is known as the “deemed export” rule — sharing controlled technology with a foreign national inside the country counts as exporting it abroad .

Anthropic determined it could not practically enforce a per-nationality restriction on a commercial product. The company’s only route to compliance was to shut down Fable 5 and Mythos 5 for everyone . On the evening of June 12, Anthropic announced the shutdown on its official X account and confirmed that access to its other Claude models was unaffected .

Anthropic’s response and the dispute

Anthropic complied with the order but did so publicly stating that the government had “got this one wrong” . The company argued that cutting off legitimate security researchers — including foreign nationals who contribute heavily to the AI workforce — would make the internet less safe rather than more . CEO Dario Amodei received the letter from Commerce Secretary Howard Lutnick directly .

The dispute highlighted a deeper contradiction. Anthropic had already been listed on a Pentagon blacklist that classified the company as too hazardous for government use, and now the Commerce Department was simultaneously restricting its most advanced models from international use .

Some reports suggest the specific trigger for the order may have been a jailbreak vulnerability in Fable 5. The Independent reported that Amazon CEO Andy Jassy raised concerns with U.S. officials about the jailbreak, which could allow the model to perform tasks beyond its intended safeguards . Anthropic pushed back, arguing that “a narrow potential jailbreak” should not justify recalling a commercial model . The government did not publicly confirm its specific security concern .

What remains unknown

The specific legal authority invoked by the administration has not been publicly detailed, though it is likely rooted in the International Emergency Economic Powers Act (IEEPA) or the Export Administration Regulations . The precise intelligence or events that prompted the June 12 action have also not been disclosed. TechCrunch noted that Anthropic’s own repeated warnings about Mythos’s danger may have “backfired,” drawing government attention that led to a more aggressive response than the company expected .

The broadness of the order raises open questions. It applied to “all foreign nationals,” with no public distinction between adversarial states and allies like the U.K., Japan, or Australia. Whether Anthropic can eventually field a U.S.-only version of Mythos-class models remains unclear. And whether other frontier labs — OpenAI, Google DeepMind, or others — will face similar export control orders for their most advanced systems is an unresolved and pressing question for the industry.

What is clear is that the shutdown represents an inflection point. A company that tried to manage risk through voluntary self-restriction was overridden by a government that judged even a hobbled public version too dangerous for international diffusion. The collision between voluntary safety and mandatory export controls has now moved from theory into practice.