Anthropic Mythos Shutdown: What Happened When the U.S. Government Killed a Trillion-Dollar AI Model

The Version That Was Publicly Released — and Then Disabled

Claude Mythos 5 was publicly launched on June 9, 2026, just three days before the shutdown order . It was the full production-grade version of the model, distinct from the earlier, more restricted Claude Mythos Preview. After the June 12 directive, Anthropic disabled Mythos 5 for all customers and users worldwide . The Mythos Preview, by contrast, remained accessible only through the closed Glasswing program and was not affected by the export control order, though the distinction was not always clear in public reporting. A Bloomberg report later confirmed that several early users of Mythos still had access to the preview after the U.S. order .

What Triggered the Shutdown?

The government told Anthropic that its directive was triggered by research on a "narrow potential jailbreak" — a method to bypass safeguards that could allow exploitation of critical software vulnerabilities . Anthropic disputed the severity of the threat, but the government’s action marked the first time it had used an export control directive to force a major AI company to pull a publicly deployed model from its own customers . Anthropic stated publicly that it believed the directive was based on a misunderstanding and said it was working to restore access as soon as possible .

Who Kept Access Through Project Glasswing?

Project Glasswing is Anthropic’s controlled program for vetted partners to access the cybersecurity-focused Claude Mythos Preview model. It expanded significantly over time:

• 12 founding organizations received access at the program’s quiet launch on April 7, 2026 .
• By May 2026, the group had grown to roughly 50 initial partners, including Apple, Google, Amazon, Microsoft, Nvidia, Broadcom, Cisco, CrowdStrike, JPMorganChase, the Linux Foundation, and Palo Alto Networks .
• By June 2, 2026, Anthropic broadened the program to approximately 200 organizations across more than 15 countries, with partners reportedly including NATO, Samsung, SK Hynix, SK Telecom, Okta, and the EU’s cybersecurity agency ENISA (pending terms) .

The Mythos Preview was made available through Amazon Bedrock as a gated research preview as part of Project Glasswing . Notably, UK banks were excluded from this expansion and instead received access to OpenAI’s rival GPT-5.5 Cyber model .

Was ENISA Denied Access?

Contrary to some early reporting, Anthropic offered the European Union Agency for Cybersecurity (ENISA) access to Mythos through Project Glasswing in early June 2026 . Bloomberg and other outlets reported that Anthropic would let ENISA join Project Glasswing, making it the first non-U.S., non-U.K. public institution to gain access . However, as of June 3, an ENISA spokesperson stated that access had been "offered but the conditions are still being negotiated" . The access was not outright denied; rather, it was held up by weeks of contentious negotiations over the terms, including usage restrictions and data-sharing rules . The situation became more uncertain after the June 12 export control directive, which raised questions about how the U.S. order might affect ENISA’s access .

The Deteriorating Relationship with the U.S. Government

The Mythos shutdown was the culmination of a broader and escalating conflict between Anthropic and the Trump administration over military AI policy.

• Pentagon blacklist and government-wide ban (February 2026): President Trump mandated that the entire U.S. government cease using Anthropic products, and the Pentagon designated Anthropic as a national security threat and supply-chain risk . This followed a contentious battle between Anthropic and the Pentagon over the company’s restrictions on military use of its AI models .
• Core dispute — domestic surveillance and autonomous weapons: Anthropic had refused to allow its AI models to be used for domestic surveillance and fully autonomous weapons systems, drawing a firm ethical line that the Pentagon and administration rejected . The company’s safety-focused constitution and acceptable-use policies explicitly prohibited these applications, which the administration viewed as hampering national security capabilities .
• Escalation to export controls: The government’s response escalated from a procurement ban to invoking export control authorities to shut down Anthropic’s flagship products entirely, an unprecedented move against a major U.S. AI company . Pentagon CTO Emil Michael stated in May 2026 that while Anthropic remained a supply-chain risk, the Mythos situation was a "separate national security concern" being addressed across the government .
• Aftermath: Following the June 12 shutdown, Anthropic’s CEO headed to Washington for meetings with the administration in an attempt to regain access to its models . As of mid-June 2026, the dispute remained unresolved, with Anthropic publicly arguing the government "got this one wrong" . The company’s official statement on X said: "We believe this is a misunderstanding and are working to restore access as soon as possible" .