US Restricts Anthropic Model Access Over Security Risk

On June 12, 2026, the US Department of Commerce, citing national security authorities, ordered Anthropic to suspend all access to Fable 5 and Mythos 5 for any foreign national, inside or outside the United States. Because Anthropic had no reliable way to verify nationality in real time, it disabled both models for all users worldwide within hours of receiving the directive, according to Anthropic's own statement. Reuters and other outlets reported the trigger was a finding by Amazon researchers that a prompting technique could get Fable 5 to identify, and in one case help exploit, a small number of already-known software vulnerabilities.

1. June 9, 2026

   Anthropic launches Fable 5 (public, heavily safeguarded) and Mythos 5 (limited to Project Glasswing cybersecurity partners).

2. June 12, 2026

   The Commerce Department directs Anthropic to suspend all foreign-national access; Anthropic disables both models worldwide.

3. June 26, 2026

   Commerce approves restoring Mythos 5 to a set of US organizations, per a letter from Secretary Howard Lutnick.

4. June 30, 2026

   Lutnick announces Commerce has fully lifted the export controls on both models.

5. July 1, 2026

   Fable 5 access is restored globally; Mythos 5 returns to Project Glasswing partners.

Anthropic says its testing found the reported bypass was not a universal jailbreak, and that comparably capable models, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, could surface the same vulnerabilities without any bypass. The company says the government provided only "verbal evidence of a potential narrow, non-universal jailbreak." As part of the resolution, Anthropic built a new safety classifier that it says blocks the specific reported technique in over 99% of cases, though at the cost of flagging more benign coding requests; Commerce's Center for AI Standards and Innovation (CAISI) tested both the prior and new safeguards and told Anthropic they are extraordinarily strong.

Reuters and ITNews reported more than 80 cybersecurity executives, including leaders from Nvidia and Adobe, signed an open letter urging the restrictions be lifted, arguing they hindered defensive vulnerability research. The Economist and other outlets framed the episode as a shift in US export-control policy from hardware toward directly gating access to deployed models, a shift that coincided with the White House's June 2, 2026 executive order on AI security and with OpenAI facing a similar government-requested limit on its GPT-5.6 rollout, according to Axios.

Anthropic, Amazon, Microsoft, Google, and other Glasswing partners are now drafting a shared framework for scoring jailbreak severity along four axes - capability gain, breadth of capability gain, ease of weaponization, and discoverability - meant to give AI developers and regulators a consistent bar for when a finding warrants a shutdown versus a patch. Anthropic also committed to give designated government evaluators pre-release access to future frontier models and to share threat intelligence in advance of publication, changes that enterprise and government customers relying on Anthropic's roadmap should track.

• •Whether the Commerce-Anthropic jailbreak-severity framework becomes a template applied to other frontier labs, including OpenAI's GPT-5.6 restrictions.
• •How the administration's August deadline, set under the June 2 executive order, for standardized AI security benchmarks plays out.
• •Whether China's push toward Mythos-class capability shapes how quickly the US eases remaining restrictions on Mythos 5 access for international Glasswing partners.

This episode is likely to be read as a case study in how ad hoc, letter-based export directives, rather than a codified statutory process, can still reach deployed AI models, despite Anthropic's own public position that such actions should follow transparent, technically grounded rules. The three-week resolution timeline, and the fact that the government ultimately accepted mitigations rather than requiring a permanent ban, suggests the administration is using export authority as a pressure lever to extract specific safeguards and information-sharing commitments rather than as a blanket bar on frontier model deployment.