US orders Anthropic to pull Fable 5 and Mythos 5

Just days after rolling out its most capable models to date, Anthropic got hit with an abrupt federal shutdown.

On June 12 at 5:21 PM ET, the U.S. government slapped Anthropic with an export control directive citing national security authorities. The order mandated the immediate suspension of all foreign national access to Fable 5 and Mythos 5, covering users abroad, non-citizens inside the U.S., and even Anthropic’s own foreign-national employees. Because isolating users by nationality on the fly is virtually impossible, Anthropic had to pull the plug entirely, disabling both models globally across all platforms. Access to the company's older models was unaffected.

The shutdown came just three days after their June 9 debut, where Anthropic had hailed Fable 5 as outperforming any model it had ever released, with Mythos 5 serving as the same core engine with specific safeguards relaxed.

A disputed jailbreak and a messy scramble

The White House order reportedly stemmed from claims that researchers found a method to jailbreak Fable 5, coaxing it into revealing software vulnerabilities usable in cyberattacks. In an unexpected twist, investor Amazon was reportedly one of the parties that flagged the potential vulnerability to officials, alongside swirling reports from Semafor that officials were spooked by unverified fears that a China-linked group had accessed Mythos.

Anthropic pushed back hard against the technical reasoning. The company stated that the government provided only verbal evidence of a narrow, non-universal bypass that involved having the model read a codebase to patch minor flaws. Anthropic insisted that these were relatively simple, known issues that existing public models like OpenAI’s GPT-5.5 could already surface without needing any jailbreak at all.

Before launch, Anthropic had spent thousands of hours red-teaming Fable alongside the U.S. and UK AI Safety Institutes. It had even implemented a strict 30-day data retention policy to track malicious prompts despite pushback from paying enterprise clients. Anthropic warned that if discovering a narrow bypass is enough to force a commercial recall, no frontier AI lab will be able to deploy new systems.

The crisis prompted rapid-fire diplomacy. Anthropic executives rushed to Washington, D.C., and CEO Dario Amodei reportedly had a private White House dinner with Donald Trump. After roughly two weeks offline, Commerce Secretary Howard Lutnick issued a letter on June 26 revising license requirements, allowing Mythos 5 back online for select organizations and clearing a path for Fable 5's return to Claude platforms, AWS, Google Cloud, and Microsoft Foundry.

Why it matters

For developers and enterprises, this was an alarming reality check on model reliability. Relying solely on the bleeding-edge frontier model can leave critical workflows stranded overnight over regulatory moves completely outside your control. It underscores why technical architecture must always maintain fallbacks to alternative models.

The broader regulatory precedent is messy. Anthropic and Dario Amodei have consistently advocated for stronger government guardrails, but this chaotic intervention was likely not what they envisioned. Enforcing a sudden global recall over verbal reports of minor capabilities sets an unpredictable precedent, especially while Washington simultaneously ponders taking equity in rival OpenAI.

We think the government's approach here was dangerously erratic. Industry consensus is that absolute, foolproof jailbreak prevention does not exist today. If every narrow vulnerability triggers an export ban and an all-hands shutdown, frontier progress grinds to a halt. It looks like proactive compliance and rigorous red-teaming are still no shield against sudden, politicized administrative strikes.