Anthropic Halts Access to Fable 5 and Mythos 5 Due to U.S. Export Control Directive
In a significant move, Anthropic has disabled all access to its artificial intelligence models, Fable 5 and Mythos 5, following a sudden export-control directive issued by the United States government. This announcement, which sent ripples through the tech industry, was made at 5:21 PM ET on June 13, 2026. The directive, citing urgent national security concerns, prohibits foreign nationals from accessing these models.
The implications of this restriction are sweeping, as it legally prevents not only public users but even foreign-national employees within Anthropic itself from interacting with the models. In an effort to comply with the newly imposed regulations, Anthropic opted to disable both Fable 5 and Mythos 5 globally, while still allowing continued operation of its earlier legacy models.
The government’s intervention is reportedly a response to the revelation of a new method capable of bypassing what has been described as Fable 5’s safety guardrails. This news of a potential "jailbreak" adds to the complexity of navigating AI safety and security protocols during an era marked by rapid technological advancement.
While the specific technical details surrounding the directive were not fully disclosed, Anthropic has indicated that the government provided verbal evidence of a limited, non-holistic jailbreak. This exploit reportedly enables manipulation of prompt inputs, compelling the model to examine a specific codebase to autonomously identify and potentially rectify software vulnerabilities.
Following an internal assessment of this technique, Anthropic concluded that the exploit had merely uncovered a small number of previously documented vulnerabilities—ones that were already known. The company argues passionately that the exploit does not provide any unique advantages or capabilities for zero-day discovery, suggesting that similar vulnerabilities could be identified through publicly available models.
Anthropic refrains from downplaying the seriousness of the situation while presenting a counter-narrative to the government’s concerns. They maintain that the rigorous security evaluations conducted prior to Fable 5’s deployment further validate the model’s resilience. This assessment included thousands of hours of stress testing performed by internal red teams, the UK AI Safety Institute (AISI), private security researchers, and relevant U.S. government entities.
These extensive evaluations suggest that Fable 5’s safety features are superior to those of previous frontier models. Testers reportedly did not uncover any universal jailbreaks capable of bypassing the model’s safeguards across a variety of cyber-attack vectors.
Recognizing the challenge of achieving perfect jailbreak resistance, Anthropic developed Fable 5 with a robust defense-in-depth architecture. This design prioritizes a series of core strategies aimed at limiting the scope of any non-universal jailbreak attempts and employing technical safeguards that deter the development of a broader jailbreak framework—essentially making it economically unfeasible.
To bolster its protective measures, Anthropic implements continuous telemetry monitoring. This proactive approach allows the company to swiftly detect any abnormal prompt behavior and act to neutralize active threats. Furthermore, a 30-day data retention policy for Mythos-class models is enforced to support ongoing threat hunting and rapid patch development.
Despite complying with the directive under substantial protest, Anthropic has expressed concerns that such an extreme recall standard could stifle innovation within the technology sector. The company argues that halting the deployment of advanced models over minor, benign responses or known vulnerabilities could set a dangerous precedent—one that could freeze future releases across major AI players.
In response to the directive, Anthropic has issued a statement asserting the necessity of governmental authority to block demonstrably unsafe AI initiatives. However, they emphasize that such actions must be based on transparent, fair, and technically sound criteria.
Amidst these challenges, Anthropic has apologized to its enterprise customers for the immediate operational disruptions. In a concerted effort to navigate this landscape, the company is actively seeking to restore access to its models. Additionally, security researchers can expect a detailed technical disclosure concerning the jailbreak mechanics within 24 hours, shedding light on the complex interplay between government oversight and technological innovation in the rapidly evolving AI landscape.
As the situation continues to develop, the implications of this export-control directive may reverberate beyond Anthropic, potentially impacting the entire artificial intelligence sector.