Anthropic shut down Claude Fable 5 and Mythos 5 for the public on June 12 after the US Commerce Department ordered the company to block access by foreign nationals. But a select group of roughly 200 organizations participating in Project Glasswing can still use Mythos Preview, an experimental version of the same model built for cybersecurity research.
The emergency export-control directive, signed by Commerce Secretary Howard Lutnick, marked the first time the federal government has directly intervened to switch off a deployed commercial AI system. Anthropic responded by deactivating both models globally, since its platform cannot verify user nationality in real time.
Project Glasswing participants including Cisco Systems, Dragos, AWS, and JPMorgan Chase have retained access to Mythos Preview, which the company uses to hunt for software vulnerabilities. Dragos CTO Jon Lavender confirmed his company still has access, and Cisco confirmed the same. The situation creates a curious loophole: the most technically capable version of the model, the one that can autonomously identify exploitable flaws in operating systems and browsers, remains online for a private network, while the public-facing version was pulled.
European cybersecurity agency ENISA, which had been invited to join Glasswing before the government's block, was informed last week it would no longer receive access. That suggests Anthropic is assessing participation on a per-organization basis against US export rules, though the company has not disclosed its criteria.
The underlying conflict traces back to July 2025, when the Pentagon awarded $200 million AI contracts to Anthropic, OpenAI, Google, and xAI. Relations fractured after US special operations forces deployed Claude during a January 2026 raid in Venezuela.
Anthropic demanded guarantees Claude would not be used for fully autonomous weapons or mass surveillance. The Pentagon refused, and Anthropic sued.
Hoping to bypass the military impasse, Anthropic launched Fable 5 and Mythos 5 to the commercial market on June 9. Three days later, the White House called at 1:00 p.m. ET warning the models posed a national security threat and demanded they be disabled within 90 minutes.
The official trigger: Amazon researchers reportedly bypassed Fable 5's safety filters, and the NSA confirmed the guardrails could be stripped. Amazon CEO Andy Jassy called Treasury Secretary Scott Bessent directly to report the finding.
High-level talks began June 15 in Washington, with Anthropic's delegation meeting Commerce Department officials. The Commerce Department has signaled willingness to let Fable 5 resume commercial operations, but export controls remain until the jailbreak vulnerabilities are resolved to the administration's satisfaction.
The White House has signaled that any future private AI model exceeding current performance thresholds must undergo government review before public deployment. For an industry built on rapid, global releases, the precedent is clear: a federal agency can now unilaterally pull a trillion-dollar product offline over a weekend.
