Fable 5 and Mythos 5 Have Been Switched Off by Government Order for 14 Days — and OpenAI Filled the Gap

An AI model went dark — not from a bug, but from a government order

Here's the deal: as of June 26, Anthropic's top models, Claude Fable 5 and Mythos 5, have been forcibly disabled for 14 days under a U.S. export-control directive. No server crash, no bug. The government ordered that the models not be made available to foreign nationals, and to comply, Anthropic pulled both models entirely — for every customer.

The trigger was surprisingly small. On June 10, an anonymous user named "Plenty the Liberator" posted a jailbreak for Fable 5 on X. Then on June 12 at 5:21pm ET, the U.S. government, citing national security authorities, issued an export-control directive: block access to Fable 5 and Mythos 5 for any foreign national — inside or outside the U.S., including Anthropic's own foreign-national employees.

Why is this big? This isn't ordinary downtime. It's the first vivid proof that an AI model can vanish overnight due to "policy risk." We'd assumed AI could be blocked for terms violations or safety issues — but the scenario where a single government directive switches off a model for customers worldwide at once felt abstract. That abstraction just became real.

So today's story: exactly what happened, why Anthropic complied yet pushed back, how rival OpenAI moved during the 14-day gap, and what it all means for those of us who use AI. Start with the cast.

The cast — Anthropic, the U.S. government, and an anonymous jailbreaker

First, Anthropic. The safety-first company behind Claude. The irony: the company that has led most loudly on "safety" had to forcibly pull its top models over a "safety concern." Anthropic complied immediately, disabling both — but in its official statement, it directly disagreed with the government's judgment. A subtle posture: comply, but do not consent.

Next, the U.S. government. It hasn't disclosed exactly which agency led, but the key is that it applied an "export-control / national-security" frame to an AI model. Export controls have mostly hit hardware like chips; here, "access to an AI model itself" became the controlled item. A signal that AI is starting to be treated like a strategic good.

Third, the anonymous jailbreaker "Plenty the Liberator." The person who posted the Fable 5 jailbreak on X. By Anthropic's account, the jailbreak is "narrow and non-universal" — it works by asking the model to read a specific codebase and fix software flaws. One person's post moved a government, and the government pulled a model used by hundreds of millions. A textbook butterfly effect.

Tie the three together: a single anonymous jailbreak post triggered a national-security directive, and that directive forced a self-described safety company to keep its own top models switched off for 14 days. That's the spine.

What actually happened

Words scatter, so here's a timeline.

When	Event
June 10	"Plenty the Liberator" posts a Fable 5 jailbreak on X
June 12, 5:21pm ET	U.S. government issues export-control directive — block all foreign-national access
June 12	Anthropic immediately disables Fable 5 and Mythos 5
June 13	CNBC, Fortune, and others report; fallout spreads
June 25	OpenAI says GPT-5.5-Cyber surpasses Mythos 5 on CyberGym at 85.6%
June 26	Day 14 offline — still no restoration timeline

Line by line. First, the scope of "block all foreign nationals" is striking. It didn't just block a specific country — it covered foreigners living inside the U.S., even Anthropic's own foreign-national staff. Precise per-user filtering was impractical, so Anthropic pulled both models for all customers. One line of order metastasized into "block everyone."

Second, Anthropic's rebuttal is the crux. It publicly stated it "disagrees that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people." It also said the government's evidence was conveyed verbally. Two top models went dark on a verbally communicated concern, without documented technical evidence — fertile ground for a fight over procedural transparency.

Third, other models are fine. The block applied only to Fable 5 and Mythos 5; access to other Claude models was unaffected. Meaning the government's concern is concentrated on "the cyber capabilities of the most powerful frontier models." The more capable the model, the bigger the regulatory target — a new grammar for AI safety policy.

What each side gains and loses

The U.S. government gains the security rationale of preemptively blocking "potential cyber misuse of a powerful frontier model" — the worry that a jailbroken model that can automatically find and fix code vulnerabilities could become an attack tool. But it loses something too: procedural opacity and a weak "verbal evidence" basis invite criticism that government intervention in AI can be arbitrary.

Anthropic's ledger is mixed. The loss is obvious — two top models earning nothing for two weeks, ceding the spot to a rival. Yet it oddly gains as well: the "comply but publicly dissent" stance positions Anthropic as a company that doesn't bow to government and demands procedural legitimacy on users' behalf. For a safety-first firm, it preserved a certain texture of trust.

The clearest beneficiary is OpenAI. Into the vacuum Mythos 5 left on cyber-security benchmarks, OpenAI announced on June 25 that GPT-5.5-Cyber scored 85.6% on CyberGym, surpassing it. At the exact moment a rival is hamstrung by regulation, OpenAI planted a flag in the same domain: "the strongest now is us." The timing is almost cruelly precise.

Past parallels — wins and failures

The closest parallel is chip export controls. The U.S. has progressively blocked exports of advanced GPUs to China on national-security grounds — the canonical case of "strategic tech is controllable," applied to hardware. The Fable/Mythos episode is the first time that logic extends to "access to a software model," meaning AI is retracing the history of chip controls step by step.

Another is the history of encryption export controls. In the 1990s, the U.S. classified strong encryption software as a "munition" and restricted exports. It never fully stopped the technology's spread, and the rules eventually relaxed. The lesson: code spreads anyway. With the jailbreak already public, skeptics ask whether pulling the model alone really removes the risk.

There are also cases where preemptive shutdowns worked — platforms that quickly disabled a feature over a safety concern and averted a bigger incident. The key is speed and recoverability: pull fast, scope the cause narrowly, restore swiftly, and you keep trust. What's worrying here, at day 14 with no restoration date, is precisely that the "restore" is nowhere in sight.

Competitor counter-plays

OpenAI has already moved — fill the gap with a cyber-security specialist model. GPT-5.5-Cyber's 85.6% on CyberGym sends the market a message: "a stable supplier unshaken by regulation." For enterprise security teams, choosing "a model that stays on" over "a model that can suddenly be switched off by regulation" is rational. Supply stability itself became a competitive edge.

Google (DeepMind) will counter more quietly. The Gemini camp can use this as a chance to emphasize "governance stability" — marketing low interruption risk and predictability rather than direct attacks, highlighting that its models carry lower sudden-shutdown risk.

And the open-source / self-hosting camp gets a surprise tailwind. Now that "a cloud API model can be switched off by a single government order" is proven, security-sensitive enterprises and institutions will reweigh the value of "a model I run myself." Demand to reduce uncontrollable external dependence may flow partly toward open-weight models.

So what actually changes

If you're an enterprise or developer on Claude's top models, you have to switch to alternatives right now. The lesson is clear: single-model dependence is a risk. Bind a core workflow to one model and your business stops the moment that model is pulled by regulation. An abstraction layer that lets you run multiple models in parallel is no longer optional — it's insurance.

If you follow AI safety and policy, this is a decisive precedent. The question "by what process, on what evidence, can a government pull an AI model?" just got asked for real. Is verbal evidence enough? What are the standards for after-the-fact review and restoration? The conclusion of this debate will frame AI governance going forward. We're watching that frame being built.

If you're a regular user, the direct impact is small — the free Claude works fine. But one thing is worth remembering: the more powerful the AI, the weaker the guarantee that it'll "always be on." Behind the convenience sits a new uncertainty — a tool that can disappear by policy.

One more layer — "verbal evidence" and the questions 14 days leave

The most uncomfortable core of this episode is "verbal evidence." Two top models went dark on a concern conveyed in words, not a documented technical analysis — a gaping procedural hole. If this becomes precedent, any model could be halted on administrative judgment alone, "without verifiable evidence." For AI companies, the unpredictability of "not knowing when or why you'll be switched off" emerges as the biggest business risk. Anthropic went out of its way to dissent publicly precisely because it grasps the weight of this precedent.

The second thing to chew on is "asymmetric harm." The jailbreak is already public on X for anyone to see, yet what got blocked is the hundreds of millions of legitimate users who used the model lawfully. If those seeking to abuse it find other routes while only well-meaning users lose access, you have to ask how much this measure actually reduced real risk — hence the criticism of "security theater." Of course, the government holds a different scale — "preventing the worst-case scenario comes first" — and the tension between these two scales is the essence of the episode.

Third is the weight of "14 days" itself. A one- or two-day emergency action is understandable, but two weeks passing with no restoration date is a different matter. Enterprise customers had to migrate workflows to other models, and a customer who has churned isn't guaranteed to return the instant the model comes back. So these 14 days are time that can produce a permanent shift in market share, not just lost revenue. OpenAI slipping precisely into that gap is no coincidence.

So this episode is a testbed that lays bare AI governance's unfinished rules, beyond a simple "regulator vs. company" clash. Many may agree the government has the authority to pull a model, but the details — on what evidence, by what process, for how long, and who decides restoration — remain blank. The ending of this Fable/Mythos case will be the first letters carved onto that blank page.

🥄 Three Things You're Probably Wondering

— So when do Fable 5 and Mythos 5 come back? Nobody knows yet. Anthropic only said it would "restore access as soon as possible," and at day 14 there's still no concrete date. It hinges on the government's judgment, so it's not something Anthropic can decide alone.

— The jailbreak is already public — does pulling just the model even help? That's exactly Anthropic's pushback: is recalling a model used by hundreds of millions proportionate to one "narrow jailbreak"? Whether it actually reduces risk or is largely symbolic is, honestly, too early to call.

— Can other companies' models get switched off this suddenly too? In theory, now yes. As the first precedent, the same frame could apply to other frontier models. Which makes "don't bet everything on one model" the most practical lesson here.

Further reading

• Statement on the US government directive to suspend access to Fable 5 and Mythos 5 — Anthropic
• Anthropic disables access to Fable 5 and Mythos 5 to comply with government directive — CNBC
• Anthropic disables Fable and Mythos AI models after U.S. government bars foreigner access — Fortune
• When a Government Pulls an AI Model: What the Suspension Means for Security Teams — Snyk
• Anthropic Disables Claude Fable 5 and Mythos 5 After US Government Order — MarkTechPost

Numbers and criteria are as of announcement and may change.