Anthropic suspends Fable 5 and Mythos 5 after a directive from the American government

Create my free watch

Anthropic announced on June 12, 2026, that it must suspend access to its models Fable 5 and Mythos 5, following an export control directive issued by the U.S. government on the same day. Officially, the order prohibits access to these two models for any foreign national - regardless of whether they are on American soil, including foreign employees of Anthropic. However, to ensure compliance, the company states it has no choice but to abruptly cut off Fable 5 and Mythos 5 for all of its clients. Access to other Anthropic models is not affected.

The directive was received on the same day at 5:21 PM (Eastern Time). Anthropic notes that the letter did not specify the exact nature of the national security concern cited. According to its understanding, the government believes it has become aware of a bypass method - a « jailbreak » - targeting Fable 5.

The bypass in question: making a model read code

Anthropic says it has examined a demonstration of the technique. This was used to identify « a small number of already known minor vulnerabilities », deemed relatively simple - and, notes the company, other public models are capable of discovering them without resorting to any bypass.

At this stage, Anthropic indicates that it has received only verbal evidence from the government regarding a narrow and non-universal potential jailbreak, essentially consisting of « asking the model to read a specific code base and fix its software flaws ». The company claims to have consulted the report it believes is the origin of the directive, and verified that the described level of capability is « widely available on other models (including OpenAI's GPT-5.5) » and used daily by security system defenders. It promises to provide more details within 24 hours.

Anthropic's defense: « among the most robust » safeguards

The company refers to the position outlined during the launch of Fable. It claims the safeguards are « so solid » that many users have complained about their excessively broad nature, particularly in cybersecurity-related uses. In the weeks preceding the launch, Fable would have been subjected to thousands of hours of red-teaming conducted with the U.S. government, the British AISI, several third-party organizations, and internal teams. These tests would have shown protections « significantly more effective than those of any deployed model so far ».

Moreover, Anthropic asserts that « no tester has, to date, found a universal jailbreak » - a bypass capable of largely disabling the safeguards. The company acknowledges, however, that perfect resistance is probably unattainable by any supplier at the current time, and that universal jailbreaks will likely eventually appear — a point it says it clearly stated at the release of Fable 5.

A strategy of « defense in depth »

In the absence of perfect resistance, Anthropic claims to be adopting a deep approach: making bypasses either narrow (non-universal jailbreaks) or very costly to produce (universal jailbreaks), and combining this with thorough monitoring to detect and quickly stop any successful attack. This is also, explains the company, why it enforces a 30-day customer data retention with Fable — a measure that has a real cost to customers, but allows it to study and fix jailbreaks. Anthropic estimates that this strategy brings the risks posed by Fable down to a level comparable to that of models already deployed in the industry.

Anthropic complies, but disputes

While applying the directive and removing access to both models, Anthropic expresses its disagreement: the discovery of « a narrow and potential jailbreak » cannot, in its view, justify recalling a commercially deployed model « used by hundreds of millions of people ». Applied across the entire industry, such a standard « would, in practice, effectively freeze all new model deployments » for all leading laboratories.

The company reminds that it publicly supports the ability of a government to block deployments deemed dangerous - but within the framework of a legal "transparent, fair, clear, and fact-based" process, principles to which it believes this action does not conform. Anthropic apologizes for the disruption caused to its customers, says it believes in a "misunderstanding," and is working to restore access as quickly as possible.

A geopolitical precedent, not just a regulatory one

Beyond the single case of Fable 5 and Mythos 5, the incident marks a deeper shift: the United States no longer controls only chips, computing capabilities, or model weights. It now claims the ability to interrupt operational access to an already commercialized model, in the name of national security.

This is a significant shift. Until now, the debate on AI sovereignty mainly focused on the upstream: GPUs, data centers, data sets, open or closed models. Here, control is exerted over the use itself. The question is no longer only about who can train a frontier model, but who has the right to query it, under what conditions, and with what nationality.

The injunction targeting foreign nationals, including when they are on American soil or working for Anthropic, introduces a new fracture in the global economy of AI. It transforms the nationality of the user, researcher, employee, or customer into a parameter for access to a cognitive infrastructure. For companies, this opens a difficult-to-contractually manage risk: a legally subscribed, technically available, and commercially deployed service may become inaccessible overnight for administrative reasons external to the client-provider relationship.

For the United States' allies, the signal is particularly sensitive. Europe, Canada, Japan, South Korea, or Australia are not necessarily the political targets of such a measure. But they are subject to its logic: in a national security regime, access to the most advanced American models may be suspended without fine distinctions between partners, competitors, and adversaries. AI sovereignty then ceases to be a defensive slogan; it becomes a question of operational continuity.

This precedent also offers an argument to powers challenging the US technological hegemony. Beijing may see it as confirmation that American frontier models are not merely cloud products, but revocable strategic capabilities. The more Washington uses access control as a power instrument, the more it encourages its rivals to accelerate their own sovereign, closed, or regional stacks.

The difficulty lies in the nature of the risk invoked. An advanced cyber capability is inherently dual-use: the same model can help a defender identify a vulnerability or an attacker exploit it. If the mere ability to read a codebase and suggest corrections becomes a sufficient reason to remove a model from the market, then the blocking threshold risks quickly encompassing all leading models. On the contrary, if states wait for too high a level of proof, they will act too late. It is precisely this in-between that remains today without a clear public doctrine.

The Fable 5 case is therefore not only a test for Anthropic. It foreshadows the political regime of frontier models: products marketed on a global scale, but subject to sovereign recall rights; private infrastructures, but treated as strategic assets; software tools, but governed by logics close to export controls and national security.

The criterion adopted in the coming weeks will be decisive. If a narrow, non-universal jailbreak, comparable to capabilities already available elsewhere, is sufficient to justify a global cutoff, the launch of frontier models will enter a new era: that of conditional, revocable, and geopolitically filtered deployments. If, on the contrary, the episode resolves as an excess of caution or an administrative misunderstanding, it will nonetheless reveal one thing: access to the most advanced models is no longer only a market question. It is now an attribute of power.

Our articles will primarily resurface in Google's « On the Front Page » section.