On June 12, AI research firm Anthropic temporarily halted access to its newly launched Claude models, Fable 5 and Mythos 5, which had only been made available three days prior. This decision was triggered by an “export control directive” issued by the US government, which restricts usage of the models to US nationals only.
Mythos represents Anthropic's most advanced "frontier" model. When it debuted in April, the company expressed concerns about its hacking capabilities, opting to limit its release to a few organizations, predominantly within the US tech sector, aimed at fixing vulnerabilities in critical digital infrastructures.
Conversely, Fable functions on the same foundational framework but includes added protections to discourage its application in cybersecurity scenarios. This model was what the public was introduced to last week, but it quickly faced shutdown.
A Tense Standoff
The relationship between Anthropic and the Trump administration has become increasingly strained since early 2025. The administration has accused Anthropic of developing “woke AI” and labelled CEO Dario Amodei an “ideological lunatic.”
Initial disputes revolved around issues of AI regulation and the policies governing semiconductor exports. Tensions heightened when Anthropic refused to allow the Pentagon access to its models for uses in domestic surveillance and autonomous weaponry.
In retaliation, the Department of Defense hinted at designating Anthropic as a “supply chain risk,” a classification that could have compelled military contractors to cut ties.
Circumventing Safeguards
While the US government has not explicitly detailed the reasoning behind last week’s directive, Anthropic speculated that authorities discovered a method—referred to as a jailbreak—capable of bypassing the safeguards in Fable, which are designed to prevent the model's powerful features from being exploited maliciously.
These protection mechanisms assess user prompts, categorizing them as either safe or unsafe before interaction with the AI model. If a request is flagged as unsafe, it redirects to a less potent version of the model for processing.
According to Anthropic, the government’s concerns were that these safeguards could be evaded, potentially allowing unauthorized users to obtain information useful for cyberattacks.
The effectiveness of these guardrails for large language models is limited, primarily relying on the model's ability to accurately interpret users' intentions. This is a complex task, especially given an active online community, dubbed the Undersphere by some observers, that is focused on finding ways around AI restrictions. Anthropic concedes that “perfect jailbreak resistance is not achievable for any current model provider.”
Anthropic suspects that the research fueling the government’s directive may have originated from Amazon engineers, who represent both a competitor and a significant investor in the company.
However, that wasn’t the only jailbreak incident. Within two days post-release of Fable, a researcher known as “Pliny the Liberator” made the full system prompt of Fable 5 public on platforms like X and GitHub. The system prompt, a concealed set of instructions that guides the model's behavior, could have implications that remain uncertain, but it has caught the attention of those in the Undersphere.
Ongoing Challenges and Uncertainties
One of the core challenges in ensuring the security of large language models like Fable is the incomplete understanding of their inner workings. Maximilian Kasy, an economist from Oxford University and a machine learning specialist, notes that these models tend to perform better than anticipated.
With billions of internal parameters and trained on massive datasets via machine learning techniques, these models seem to successfully generalize far more effectively than would be expected. Kasy likens the current state of AI development to alchemy—driven by experimentation rather than grounded in systematic scientific theory.
Consequently, even the creators have limited visibility into the models' behavior.
Difficulties in Regulation
The lack of transparency within this technology makes regulation a daunting task. Governments often lack the necessary access to data, infrastructure, and expertise to assess proprietary frontier models effectively. This challenge has been echoed in a recent executive order from the US administration on AI security, issued two weeks ago. It reflects a shift from a previously hands-off approach to a demand for developers to present their models for review prior to public release.
This requirement indicates an implicit recognition by the administration that it cannot fully trust companies to provide a comprehensive assessment of their models’ potential capabilities and the risks of misuse. Public sentiment is further complicated by findings from a global survey conducted last year, which indicated that individuals are more than twice as likely to express concern over AI than excitement.
Looking Ahead on AI Safety
While AI is undeniably a highly discussed and promising technology, its unpredictable and potent nature raises significant concerns. This dangerous combination necessitates more than just regulatory efforts, as the pace of technological advancements often outstrips the capacity for effective oversight.
Moreover, current guardrails may be insufficient due to the likelihood of circumvention. A governance framework is essential—one that is capable of anticipating and addressing potential failures.
Such a framework needs to be international, collaborative, and built on mutual trust. Unfortunately, the present administration in the US has shown limited ability to foster such an environment.
