Anthropic blocks Fable 5 and Mythos 5 within days of release after US export control directive

Anthropic blocks Fable 5 and Mythos 5 within days of release after US export control directive

Anthropic received a US Commerce Department directive on June 12, 2026, ordering it to block access to its newly released Fable 5 and Mythos 5 models. The two models launched on June 9 — Fable 5 as the first publicly available Mythos-tier model, and Mythos 5 as a reduced-safety version restricted to Project Glasswing participants. The directive, delivered via letter from Commerce Secretary Howard Lutnick to CEO Dario Amodei, requires that all foreign nationals — including foreign employees at Anthropic — be barred from accessing the models. Anthropic responded by shutting down access for all customers rather than attempting selective foreign-national blocking, citing the impracticality of doing so at scale.

Why it matters: This event marks an unprecedented case of the US government compelling a frontier AI lab to retroactively remove a commercially deployed model from the market under export control authority. The grounding force is not model capability but national security concern — specifically, a claimed jailbreak vulnerability in Mythos 5. Anthropic disputes the severity, stating the identified jailbreak is narrow and non-general, comparable to capabilities already present in models like GPT-5.5. The incident creates a new regulatory precedent: if the government can order takedowns for narrow vulnerabilities, it effectively gains veto power over any frontier model release, reshaping the release-approval dynamics between labs and regulators.

Grounded expert take: This is a structural force shift masquerading as a single incident. The export control mechanism has been used in novel territory — not to block chip flows to China, but to halt domestic model distribution to all foreign users globally. The Commerce Department argued the model posed national security risks in foreign hands, a rationale that could be applied retroactively to any frontier model with cyber or biosecurity relevant capabilities. Anthropic’s position — that this standard would halt virtually all new frontier model deployments if applied consistently — is not self-serving spin; it exposes the unresolved open debate about how export controls intersect with commercial AI release cycles. The practical outcome is that frontier labs now face a new binary risk: a model can be fully deployed for days, then zero-day revoked by fiat. Enterprise customers who built workflows on Fable 5 are left stranded — a contingency that may accelerate demand for on-premise, government-approved model variants.

#Anthropic #ExportControls #FrontierModels #AIGovernance #ModelTakedown #NationalSecurity