A decision by the U.S. government was enough to shut down, from one day to the next, the most powerful AI model available on the market. An export directive. The Fable 5 case makes it clear that relying on a single AI provider is not a technical detail: it is a business risk that no contract warns about.

On June 9, Anthropic presented Claude Fable 5 and Claude Mythos 5, the two most advanced AI models the company has put on the market so far. On June 17, both were shut down for all non-U.S. users, by order of the U.S. government. In light of this situation, the question arises: what happens when critical infrastructure depends on a foreign geopolitical decision.

A model that admits its own limits

Fable 5 arrived with a unique feature: Anthropic admitted that the model needs a safety brake to operate safely. When it detects that a query could be used for something dangerous, such as a cyberattack or the development of biological weapons, it automatically lowers its response level and delegates that query to a more limited version. The company estimated that this happens in less than 5% of conversations, although early users noted that the brake was activated more often than expected, even on queries that were not risky at all.

The version without this brake, Mythos 5, was not released to the public. It remained reserved for a small group of organizations working in cyberdefense and critical infrastructure, within a program Anthropic calls Project Glasswing.

In an independent test where models must detect and fix real errors in software code, Fable 5 resolved 8 out of 10 cases. GPT-5.5, from OpenAI, resolved fewer than 6. Gemini 3.1 Pro, from Google, also fell short. In the most difficult tasks, the advantage of Fable 5 tripled.

One day for the first jailbreak

On June 10, a researcher identifying as Pliny the Liberator posted on the social network X that they had managed to bypass the model's security system. Their method combined three techniques: using symbols the system did not recognize correctly, disguising queries as academic exercises, and breaking a risky query into loose parts that appeared innocent individually. By combining the responses, they reconstructed information that the model should have blocked.

Anthropic reviewed the report and clarified that it was a localized failure, not a general vulnerability of the model. The company added that this same type of vulnerability can also be found in other AI models available on the market.

The directive that shut down both models

On June 12, the U.S. government notified Anthropic of an export control directive. The order forces the suspension of access to Fable 5 and Mythos 5 for any non-U.S. citizen, inside or outside the country, with no exception for the company's own employees.

Anthropic explained that it could not segment the block between resident and non-resident users, and had to disable both models completely. The rest of its product line, including Claude Opus and Sonnet, was not affected.

The company complied with the order, but challenged it. According to its official statement, the government did not provide evidence of a universal jailbreak, only a localized one, similar to asking the model to review code and fix bugs. Anthropic compared this level of capability to that of other models used daily by defensive cybersecurity teams, and warned that applying this standard systematically could stall the deployment of any frontier model in the industry.

The risk that does not depend on technology

Franco Scapin, Co-Founder and CAIO of Rocbird, argues that the technical details of the case are secondary to the underlying lesson. "The most important thing to me is not the specific details of this particular situation, but that every company relying on a single AI provider just got a very clear reminder of platform risk," he states.

"If AI is becoming the core of your business, having a multi-model setup, or at least the flexibility to switch providers without a major rebuild, is no longer really optional," says Scapin. "Otherwise, you are a guest at someone else's table, and someone else's geopolitics could become your point of failure overnight."

The episode is built on a simple fact. An administrative decision, completely unrelated to any major technical failure, was enough to shut down the most powerful AI model on the market overnight. There was no massive security breach or data leak. There was a government directive.

Scapin anticipates that the case will push more companies toward open source and to take the concept of AI sovereignty seriously: the ability of an organization, or a country, to avoid relying on the unilateral decision of a foreign provider to keep its critical infrastructure running.

What this implies for companies already using AI in production

The Fable 5 case shifts the problem from the technical arena to management. No company decides when a government issues an export directive. It does decide how its operation is positioned when that occurs.

Designing the architecture so that no provider is a single point of failure involves concrete work: evaluating open source alternatives, keeping the integration decoupled enough to migrate without rewriting it entirely, and treating technological dependence as a risk variable in business planning, rather than as a stack detail that you solve once and forget.