Anthropic Pulls Fable 5 and Mythos 5 Offline After Commerce Directive on Export Controls

The directive arrived Friday afternoon without stated national security findings or statutory process. Anthropic described the models as already restricted, with Mythos 5 limited to vetted users due to prior cybersecurity concerns. No CVE or active exploit data was referenced in the action, yet the models were taken fully offline rather than segmented by access tier.

Procurement records and prior Commerce actions on advanced semiconductors show the same pattern: sudden extraterritorial reach applied to dual-use technology before formal rulemaking. The executive order issued ten days earlier explicitly labeled participation voluntary and required a one-month risk assessment window; the Friday directive bypassed both. Technical evidence of model weights or inference endpoints leaving U.S. jurisdiction remains absent from public filings.

This exposes the gap between announced policy and operational enforcement. Developers now face ad-hoc blocks that can be issued without published criteria or appeal timeline, raising compliance costs and fragmenting model availability. Next steps hinge on whether Commerce publishes the specific technical thresholds triggering the block or if Anthropic restores access through a narrow domestic-only API.

Independent verification of any data transfer logs or foreign inference attempts has not been released. The incident aligns with documented tightening of AI compute and model controls rather than isolated safety concerns.