US orders Anthropic to take Claude Fable 5 and Mythos 5 offline worldwide amid security fears
The U.S. government has issued an emergency export control order compelling Anthropic to shut down access to its two newest and most powerful AI models, Claude Fable 5 and Claude Mythos 5, only days after their launch.
In a sweeping directive delivered Friday, regulators told Anthropic to immediately suspend access to both models for all foreign nationals, citing unspecified national security concerns. The order applies regardless of physical location: non‑U.S. persons are barred from using the systems whether they are inside the United States or abroad.
Because Anthropic cannot reliably distinguish every user’s legal status and nationality in real time, the company has effectively been forced to pull the models for everyone to ensure full compliance. Both Fable 5 and Mythos 5 have been disabled across Anthropic’s entire customer base.
Government cites jailbreaking risk; Anthropic calls it overreach
According to the directive, officials believe they have identified a method for bypassing – or “jailbreaking” – the safety systems on the publicly available Fable 5 model. Such a technique could, in theory, allow users to circumvent guardrails and compel the model to produce restricted or harmful outputs that it was designed to block.
The letter reportedly offered no detailed technical evidence of the vulnerability and did not spell out precisely how the exploit works. Nor did it elaborate on what specific national security harm officials believe could arise, beyond the generic risk that powerful AI models might be misused once their controls are weakened.
Anthropic pushed back against the scope of the move, characterizing the decision as an overreach. The company has argued that the type of vulnerability cited by the government is not unique to Fable 5 or Mythos 5 but is already widespread across the advanced AI sector. In Anthropic’s view, almost every frontier model is engaged in an ongoing cat‑and‑mouse game with jailbreak attempts, and singling out two newly launched systems does not meaningfully reduce overall risk.
Why Mythos 5 is also in the crosshairs
Although the government’s concern centers on a jailbreak technique discovered in Fable 5, the order also covers Claude Mythos 5, Anthropic’s even more capable model in the same family. Officials appear to be acting on a precautionary logic: if one member of the model line can be compromised, more advanced variants could be vulnerable in similar or more dangerous ways.
Even without a specific, documented exploit for Mythos 5, policymakers are treating both models as part of a single risk surface. The fear is that any path that allows users to defeat alignment and safety layers in one system may generalize to others, especially those with greater reasoning, coding, or optimization capabilities.
A rare use of emergency export control powers
What makes this episode especially notable is the legal tool the U.S. government chose. Rather than going through lengthy, open rulemaking, officials invoked emergency export control authorities traditionally used for sensitive dual‑use technologies – for instance, advanced semiconductors, encryption systems, or surveillance hardware.
By framing Fable 5 and Mythos 5 as subject to export controls, the government has:
– Prohibited Anthropic from “exporting” access to the models to foreign nationals, even if they are physically in the United States.
– Placed the burden of enforcement squarely on the company, forcing it to control who can use its systems and how.
– Signaled that cutting‑edge AI models are being treated less like consumer software and more like strategic technologies with geopolitical implications.
This is a rapid escalation in the regulatory treatment of frontier AI, moving from abstract policy debates to direct intervention in product availability.
Anthropic’s dilemma: compliance vs. global user base
For Anthropic, the directive creates an immediate operational and commercial problem. The company’s customer base is global, and its own workforce includes foreign nationals. Implementing a selective, nationality‑based access cut‑off at short notice is extremely complex, fraught with legal risk, and technically brittle.
Faced with the possibility of violating U.S. export law, Anthropic’s safest short‑term option was blunt: shut down both models entirely. That decision:
– Eliminates the risk of accidental access by restricted users.
– Preserves the company’s ability to negotiate, appeal, or seek clarification without compounding potential violations.
– Comes at the cost of momentum, user trust, and competitive positioning in an intensely contested AI race.
In effect, a national security directive aimed at foreign access to powerful AI has temporarily removed those models from everyone, including American companies and developers trying to build on top of them.
A familiar weakness in AI safety – but now with legal consequences
The heart of the government’s concern – a jailbreak that weakens safety – is not new. AI researchers, red‑teamers, and hobbyists have been probing large language models for years, discovering prompts and techniques that can coax systems into ignoring their guardrails.
What is new is that such a vulnerability is now the trigger for an emergency government action. This suggests a turning point:
– Incremental safety gaps that might once have been treated as a technical issue between companies and their users are becoming matters of national policy.
– Governments are increasingly willing to view alignment failures not just as product bugs, but as security liabilities.
– Companies releasing frontier models may now need to anticipate that a widely publicized jailbreak could result in abrupt regulatory intervention.
Anthropic’s contention that “everyone has this problem” is likely accurate in a technical sense. But regulators may respond that ubiquity is precisely why they must act before such weaknesses are exploited at scale.
What “national security” might mean in this context
The directive does not spell out the exact harms officials fear, but several themes are shaping how governments think about advanced AI models:
– Scalable misuse: Even if a model cannot autonomously act in the real world, it can produce instructions, analyses, or tools that dramatically lower the barrier for bad actors.
– Model proliferation: Once a powerful system is broadly accessible, its behaviors can be studied, cloned, or distilled into less regulated copies.
– Strategic advantage: Leading AI models are seen as part of a broader technological competition; uncontrolled diffusion could narrow the gap between countries’ capabilities.
Without over‑speculating, the direction of travel is clear: the most capable models are being treated less as neutral productivity tools and more as strategic assets subject to security‑driven controls.
A preview of the regulatory future for frontier AI
The Anthropic case is likely to become a reference point in future AI policy debates. Several trends are crystallizing:
1. Model‑level controls, not just use‑case rules
Policymakers are no longer content to regulate only how AI is used. They are increasingly willing to restrict who can access certain models at all, depending on scale, capability, or perceived risk.
2. Export control logic applied to software
The decision to treat AI models like export‑controlled goods puts them in the same conceptual category as advanced chips. This could lead to licensing regimes, registration requirements, or mandatory reporting for high‑end systems.
3. Rapid, discretionary action
Emergency directives bypass the slow consensus‑building of normal regulation. Companies operating at the frontier must plan for sudden, disruptive interventions when governments believe risks have crossed a threshold.
4. Nationality as a gating factor
Basing access on citizenship or residency raises technical, ethical, and legal questions. Yet this directive shows that governments are willing to use nationality as a blunt instrument to manage perceived strategic risk.
How this affects developers, enterprises, and the broader AI ecosystem
In the near term, users who had begun experimenting with Fable 5 or Mythos 5 will have to revert to earlier Claude versions or competitor models. For businesses, this disruption highlights several practical lessons:
– Avoid single‑model dependence: Relying entirely on one frontier model now carries regulatory as well as technical risk. Multi‑model strategies and abstraction layers become more important.
– Track regulatory exposure: Enterprises in sensitive sectors may need to treat AI model selection like any other compliance‑relevant decision, especially when dealing with foreign offices or staff.
– Expect volatility in the frontier: The most advanced, most capable systems are also the most likely to face abrupt policy shocks. For long‑term projects, stability may matter as much as raw performance.
For the wider ecosystem, the message is unmistakable: the era when AI labs could freely launch their most powerful models worldwide and “iterate in public” is ending. Frontier AI is now firmly within the orbit of national strategy.
Anthropic’s next moves – and what to watch for
Anthropic will likely pursue several parallel tracks in response:
– Technical hardening: Intensified work on safety alignment, adversarial training, and automated detection of jailbreak attempts.
– Regulatory engagement: Negotiations with U.S. authorities to define conditions under which the models might be re‑enabled, at least for restricted user groups.
– Transparency and assurances: Providing more detailed documentation of safety evaluations, red‑teaming results, and mitigations to rebuild confidence among both regulators and customers.
Policy‑watchers will be looking for:
– Whether the government narrows the order to specific countries, organizations, or risk profiles.
– Whether a clear standard emerges for what level or type of jailbreak triggers intervention.
– How other AI labs respond – whether by slowing public releases, geo‑fencing models, or pre‑emptively restricting access.
A new baseline: powerful AI as regulated critical tech
The forced withdrawal of Claude Fable 5 and Claude Mythos 5 sets a new practical baseline for the AI industry. It demonstrates that:
– Governments are prepared to act quickly and decisively against specific models, not only through broad, future‑oriented laws.
– Safety vulnerabilities that were once seen as inevitable growing pains of a new technology can now carry immediate legal and commercial consequences.
– Advanced AI capabilities are, in the eyes of state actors, converging with other forms of critical, strategically sensitive technology.
As Anthropic works to satisfy regulators and restore access to its latest systems, every major AI developer now has a clearer picture of the new environment: model design, deployment, and safety are no longer just engineering problems or market decisions. They are intertwined with export controls, national security assessments, and the geopolitics of intelligence and computation.
In that world, releasing a frontier AI model is no longer just a product launch. It is, implicitly, a negotiation with the state.