Most requirement documents describe who may use a system in terms of roles and regions. The June 12 directive to Anthropic describes Fable 5 and Mythos 5 in terms of foreign nationals — a category that ignores geography, employment status, and the way API keys are actually issued. Three days after Anthropic opened Fable 5 to subscription customers, both Mythos-class models went dark for everyone. US-based teams that never touched the restricted tier included.

That sequence reads like a vulnerability recall. The public record says otherwise: an export-control order whose compliance geometry forced a global kill switch, sitting on top of a jailbreak dispute Anthropic calls narrow, verbal, and not Mythos-specific.

Hard-coded claude-fable-5 last week? Your incident is policy latency — not corrupted weights.

What broke on Friday — and what did not

Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9, 2026. Fable 5 was the safeguarded general-release face of a new capability tier Anthropic calls Mythos-class — models above Opus in raw skill, priced at $10 per million input tokens and $50 per million output tokens. Subscription plans briefly included Fable at no extra cost through June 22. Teams that migrated agent workflows over the weekend were routing to a model that stopped existing by Friday evening.

The directive arrived at 5:21 p.m. ET on June 12. Anthropic's June 12 update on its launch page is blunt: access to both models is suspended while the company works to restore service. Developer notices describe the operational fallout — new sessions fall back to a user's default model or Opus 4.8, in-flight Fable sessions error out, and Platform requests targeting claude-fable-5 fail until integrators reroute.

What did not break is the rest of Claude. Opus 4.8 remains fully available across surfaces the reporting covers. If your integration never pointed at Fable or Mythos, your outage is indirect — a roadmap item paused, not a platform blackout. The selective targeting matters when you interpret the government's concern: this is not "frontier AI is too capable to sell," at least not yet. It is these two SKUs, now.

Compliance geometry — why "inside the US" still fails the order

Export control language sounds geographic until you read the operative restriction. Reporting consistently describes a directive citing national security authorities that bars access by any foreign national, whether that person is inside or outside the United States, including Anthropic's own foreign-national employees. That wording is the hinge.

Picture a delivery lead in Boston on an H-1B whose team adopted Fable 5 for a long-horizon code migration on Tuesday. Under the letter of the order, that engineer is a foreign national using a covered model inside the US. Anthropic's SaaS stack, like most API products, authenticates accounts and organizations. Not immigration status per request. There is no practical way to allow "US persons only" sessions on Fable without building nationality attestation into the hot path, and Anthropic has said it cannot filter foreign nationals from US users in real time.

So compliance became binary: disable Fable 5 and Mythos 5 for all customers, or risk serving someone the order excludes.

The worldwide outage is compliance geometry — nationality-scoped law meeting borderless identity — not proof that every country was judged equally dangerous.

Global kill switch. Nationality-scoped order.

That asymmetry is what most coverage skips.

Fable and Mythos — same weights, different policy surface

Anthropic's launch materials draw a line analysts should keep straight. Fable 5 and Mythos 5 share the same underlying Mythos-class model; the names differ because the safeguards differ. Fable ships with classifiers that detect potential misuse — jailbreak attempts, cyber-offense tasks, sensitive biology and chemistry queries, distillation patterns — and route flagged traffic to Opus 4.8 instead of answering from Fable directly. Early data Anthropic published claims more than 95% of Fable sessions never trigger fallback; for those sessions, performance is effectively the same as Mythos 5.

Mythos 5 is the other face of the same capability: cyber safeguards lifted for vetted defenders under Project Glasswing, with plans to expand through trusted-access programs consulted on with the US government. That split is deliberate product policy — speed a safeguarded general release while reserving reduced-friction access for screened partners. It also creates two export-control surfaces on one weights file: a mass-market SKU and a restricted SKU whose whole purpose is higher cyber uplift for approved users.

When the directive names both models, it is not treating them as unrelated products. It is targeting the Mythos-class tier Anthropic just elevated from preview to broad availability — the tier whose launch post already acknowledged substantial cyber and dual-use risk and documented red-teaming, bug bounties, and 30-day retention for safety monitoring on Mythos-class traffic.

The jailbreak trigger — security language, disputed evidence

Secondary reporting and Anthropic's own statements converge on a trigger behind the directive: a reported path to jailbreak Fable 5. Anthropic's public response is where the "security recall" framing breaks down. The company says the government provided only verbal evidence of a potential narrow, non-universal jailbreak — described in reporting as essentially asking the model to read a codebase and fix software flaws. Anthropic reviewed a demo and characterized the findings as minor bugs, the kind other publicly available models surface without a bypass.

The Verge quotes Anthropic going further: no written specifics of the national security concern, no disclosure of a non-universal jailbreak that led to a harmful result, and findings that provide no Mythos-specific uplift relative to models such as GPT 5.5. Whether you find that credible or performative, it defines the dispute. Anthropic is complying with a legal directive while arguing the evidentiary standard would halt frontier deployments industry-wide if applied uniformly.

That is not the same as claiming Fable's safeguards are perfect. Anthropic's own launch post noted progress toward a universal jailbreak in a brief UK AISI testing window — while also reporting no universal jailbreaks from a 1000+ hour external bug bounty.

Security teams ask whether the bypass scales. Regulators ask who may receive uplift if it does.

The June 12 directive answers the second question first — with export-control tools, not a CVE-shaped recall path.

What delivery teams should do while access is suspended

Operationally, the move is reroute, not panic. Point new agent sessions and API calls at Opus 4.8 or your prior default until Anthropic restores Fable or publishes a segmented access program. Treat any workflow that hard-coded claude-fable-5 as a configuration debt item — model IDs on frontier tiers are policy-sensitive assets now, not just performance knobs.

Strategically, read selective targeting as signal. Opus 4.8 still serves the mass market; only Mythos-class SKUs were pulled. That pattern fits a specific-trigger story — export control plus a disputed jailbreak path on the newest tier — better than it fits "all frontier capability is deemed uncontrollable." It does not mean your Opus integrations are permanently safe from similar orders. It means the June 12 action was model-scoped, which is how export regimes often behave when the concern is transfer of capability to excluded persons rather than a vendor-wide product failure.

Watch Anthropic's promised follow-up within 24 hours of the directive for restoration terms. Restoration, if it comes, will likely involve access controls tighter than "any paying customer" — the trusted-access shape Mythos was already moving toward — not a simple flip of the feature flag.

The order was about who may hold the capability.

The outage was about who the vendor could prove at the door.