The United States government has taken the unprecedented step of shutting down Anthropic's latest artificial intelligence models, Claude Fable 5 and Mythos 5, just days after their release. On June 12, the AI lab suspended access to these models following an "export control directive" that prohibits use by anyone who is not a US national.
Background on the Models
Mythos 5 is Anthropic's most powerful frontier model. When first announced in April, the company acknowledged it was too proficient at hacking for immediate public release. Instead, it was made available only to select organizations, mostly US tech corporations, to help patch vulnerabilities in critical digital systems. Fable 5 is essentially the same model but with additional safeguards designed to prevent its use for cybersecurity purposes. It was released to the public on June 9 and shut down almost immediately.
Conflict with the Trump Administration
Since early 2025, tensions have been escalating between Anthropic and the Trump administration. The administration has accused Anthropic of producing "woke AI" and labeled CEO Dario Amodei an "ideological lunatic." Early disagreements revolved around AI regulation and semiconductor export policies. The conflict intensified when Anthropic refused to allow the Pentagon to use its models for domestic surveillance and fully autonomous weapons systems. In response, the Department of Defense threatened to designate Anthropic a "supply chain risk," which would force military contractors to sever ties with the company.
Jailbreak Concerns
The US government has not publicly stated the reason for the directive, but Anthropic believes it stemmed from the discovery of a jailbreak—a method to bypass Fable's safeguards and access its most powerful features for malicious purposes. These safeguards classify user requests as safe or unsafe before passing them to the AI model, redirecting unsafe requests to a less powerful model. The government's concern was that the safeguards could be circumvented to extract information useful for cyberattacks.
Guardrails for large language models are not foolproof. They rely heavily on the model's ability to interpret user intentions. Additionally, a vast online community known as the Undersphere actively works to circumvent AI guardrails. Anthropic acknowledges that "perfect jailbreak resistance is not achievable for any current model provider."
Anthropic claims that the research behind the government directive appears to have been produced by engineers at Amazon, which is both a rival and a significant investor. However, this was not the only relevant jailbreak. Within 48 hours of Fable's release, a researcher using the pseudonym "Pliny the Liberator" published what they identified as Fable 5's full system prompt to X and GitHub. The system prompt is a hidden set of instructions that helps determine an AI model's behavior. While it is unclear how knowledge of the system prompt could be used in practice, it has drawn significant attention in the Undersphere.
The Opacity of AI Systems
A fundamental challenge in securing large language models like Fable is that we do not fully understand how they work. According to Oxford University economist and machine learning expert Maximilian Kasy, these models perform much better than they "should." Large language models have billions of internal parameters and are trained on enormous datasets using machine learning. Kasy notes that we would expect such systems to be "overfitted"—good at reproducing patterns in training data but poor at generalizing to new situations. However, modern systems like Claude and ChatGPT do appear to generalize effectively. Kasy compares modern AI development to alchemy: successful through trial and error, but not yet grounded in systematic theory. As a result, the behavior of AI models remains partly opaque even to their creators.
Regulatory Challenges
The opacity of AI technology is a key reason it is so difficult to regulate. Governments lack independent access to the data, infrastructure, and expertise needed to evaluate proprietary frontier models. The US administration's recent executive order on AI security, published two weeks ago, reflects this realization. Having recognized the power of frontier AI models, the administration has shifted from an initial hands-off approach to requesting that developers share their models for review before release. This demand implicitly acknowledges that the administration does not trust companies to fully and comprehensively evaluate what their models can do and how they might be misused.
The public sees even less, and the consequences are measurable: a survey across 25 countries last year found that people are, on balance, more than twice as concerned about AI as they are excited about it.
The Future of AI Safety
AI is a hugely hyped technology, but it is also extremely powerful and unpredictable. This combination is understandably very dangerous. We cannot rely solely on regulations, as technology will develop faster than they can adapt. Nor can we rely on guardrails, as they will be bypassed. What is needed is a governance framework built for that eventuality—one that can predict and address the consequences of failure. Such a framework must be global, participatory, and founded on reciprocal trust. These are qualities the current US administration has shown little capacity to foster.
