The US Commerce Department issued an export control directive on 12 June ordering Anthropic to pull Claude Fable 5 and Mythos 5 offline globally — an unprecedented government action that immediately disrupted enterprise AI workflows worldwide.
The directive followed two converging events. On 10 June, a jailbreaker known as 'Pliny the Liberator' posted a viral demonstration of a 'pack hunt' — a coordinated multi-agent jailbreak using Unicode substitution and decomposition techniques that broke harmful requests into benign sub-questions. Separately, Amazon researchers submitted findings showing the models could complete a standard defensive security workflow: reviewing vulnerable code and then fixing it.
David Sacks, co-chair of the President's Council of Advisers on Science and Technology, later disclosed that the administration offered Anthropic a choice before issuing the directive: fix the jailbreak or voluntarily de-deploy Fable 5. CEO Dario Amodei refused both options, treating the administration's characterisation of risk as technically unfounded. Anthropic argued that the jailbreak was technique-specific rather than a universal vulnerability, and that the same defensive security capabilities exist in other publicly available models.
The shutdown was triggered specifically by SK Telecom's access patterns, which the White House identified as having suspected ties to China. The two-step sequence saw Anthropic first revoke SK Telecom's access, followed by the broader global ban after the Amazon vulnerability report.
The impact was immediate. Current sessions routing to Fable 5 ended in errors, with new queries automatically routing to Opus 4.8. Enterprise customers who had optimised workflows for Fable 5 faced significant productivity disruption.
For context engineers, the Fable 5 shutdown establishes a precedent with far-reaching implications: frontier AI models are now subject to government recall via export control mechanisms, regardless of whether the company agrees with the risk assessment.
