Anthropic has suspended access to its newly released Claude Mythos 5 and Claude Fable 5 models following a direct export control order from the US government, underscoring the growing tensions between rapid AI advancement and national security concerns. The decision, announced late on June 12, 2026, came just days after the models’ launch and reflects how quickly regulatory oversight can intervene in the development of frontier AI systems.
The background involves a pair of specialized models built on earlier security-focused research. Claude Mythos Preview, developed under a limited-access program called Project Glasswing, had shown promise in helping partners identify and patch software vulnerabilities. Participants, including Mozilla, reported resolving hundreds of issues with its assistance. When Anthropic rolled out the full Mythos 5 and the more broadly intended Fable 5 versions earlier this week, it maintained restrictions on Mythos while making Fable available to the public with additional safety guardrails. Those safeguards quickly became a point of debate, as the company adjusted them in response to user feedback on usability versus risk.
The government’s intervention arrived abruptly. According to reports, Commerce Secretary Howard Lutnick issued a directive subjecting both models to export controls, effectively barring access by foreign nationals both inside and outside the United States. This includes Anthropic employees who are not US citizens. The trigger appears to have been a claimed jailbreak of Mythos 5, which raised alarms about potential misuse despite the vulnerabilities demonstrated being relatively minor and replicable in other leading models, including OpenAI’s offerings. Anthropic received the order at 5:21 PM ET and moved swiftly to comply by disabling the models entirely for all users, while emphasizing that its other Claude models remain available.
In its statement, Anthropic expressed the view that the action stemmed from a misunderstanding. The company noted that the jailbreak technique exposed only previously known, straightforward issues already discoverable by comparable public models. It highlighted the extensive safeguards implemented at launch and signaled plans to provide more details within 24 hours while working to restore access. For users, the immediate effects include errors in existing Fable 5 sessions and the need to switch integrations to alternative models like Opus 4.8 or defaults.
This episode highlights persistent challenges in AI governance. Export controls, traditionally applied to hardware and sensitive technologies, are now extending into software models with increasing frequency. The move fits a pattern of US efforts to manage perceived risks from advanced AI, even as critics argue such restrictions may slow domestic innovation without meaningfully addressing global proliferation. Jailbreaking remains a common cat-and-mouse dynamic across the industry, suggesting that no single model holds a monopoly on concerning capabilities. At the same time, the speed of the response raises questions about the balance between precaution and evidence-based policy, especially when the cited vulnerabilities do not appear unique.
For developers and researchers relying on these tools, the sudden cutoff disrupts workflows and erodes confidence in platform stability. It also illustrates the complexities of operating in a field where commercial ambitions intersect with government priorities. Anthropic’s emphasis on compliance while gently pushing back indicates a desire to maintain constructive dialogue, yet the incident may prompt broader caution among AI labs about rapid releases.
As AI capabilities continue to evolve, incidents like this one reveal the limitations of both corporate safeguards and regulatory speed. They serve as a reminder that meaningful progress requires not only technical excellence but also clearer frameworks for assessing and mitigating real risks without unnecessary setbacks. The coming days may clarify whether access can be restored or if further adjustments will be needed to satisfy oversight requirements.
