The United States has ordered the immediate suspension of access to Anthropic’s advanced AI models Fable 5 and Mythos 5, citing national security concerns under export control authorities, in a move that forced the company to disable global access to the systems.
The move comes less than a week after the models were publicly released on Tuesday, June 9, 2026, making it one of the fastest regulatory interventions on a frontier AI system to date.
According to Anthropic, the directive requires the company to block all foreign nationals from accessing the models, including foreign employees working at the firm, regardless of their physical location.
The order effectively resulted in a full shutdown of the two models for all users worldwide to ensure compliance.
The government did not publicly disclose the specific national security rationale behind the decision.
However, Anthropic said it was informed that authorities had identified a potential method of bypassing the models’ safeguards, commonly referred to as a “jailbreak,” which could allow users to access restricted capabilities.
Jailbreak concerns
In its internal assessment, Anthropic said it reviewed a demonstration of the alleged jailbreak technique and found that it revealed only “minor, previously known vulnerabilities” that were not unique to the Fable system.
The company further stated that similar outputs could reportedly be produced by other widely available AI models without requiring any bypass.
Anthropic emphasised that extensive pre-release testing of Fable 5 included thousands of hours of red-teaming in collaboration with US and UK AI safety institutes, private evaluators, and internal security teams.
According to the company, no universal jailbreak, defined as a method capable of broadly bypassing safeguards, was discovered before launch.
The firm has long maintained that while safeguards can significantly reduce misuse risks, “perfect jailbreak resistance is not currently possible” for any AI system, and that adversarial methods are expected to evolve.
Why American government acted
The export control directive highlights growing US concerns over frontier AI systems and their potential misuse in cybersecurity-related contexts.
Officials reportedly believed the identified jailbreak method could be used to access sensitive technical information, although Anthropic says it has not been provided with detailed evidence of any harmful real-world outcome.
The company also stated that the government shared only limited verbal information and referenced a narrow scenario involving software debugging tasks, such as identifying vulnerabilities in codebases.
Anthropic, however, disputed the severity of the finding, arguing that such capabilities are already widely available in existing AI systems used by cybersecurity professionals and defenders.
A key element of the directive is its restriction on foreign national access, including employees of AI companies operating in the United States.
This reflects a broader US export control approach that treats advanced AI systems similarly to sensitive dual-use technologies such as semiconductors or encryption systems.
While the government has not publicly explained its reasoning in detail, such restrictions are typically justified under concerns that advanced models could be misused by foreign actors for cyber operations, intelligence gathering, or other security-sensitive applications.
Anthropic complied with the order but questioned its scope, arguing that a narrow vulnerability discovery should not justify the recall of a commercially deployed model used at scale.
Allegations remain unverified
The company statement references internal and government-reviewed reports on a potential jailbreak method. However, claims circulating possible foreign exploitation, including suggestions of state-linked actors or “leaks”, have not been independently verified and were not confirmed in Anthropic’s disclosure.
Up to now, the Eastleigh Voice has confirmed, no evidence has been publicly presented linking the incident to any specific country, organisation, or coordinated external breach.
However, this incident highlights growing tension between rapid AI deployment and national security regulation.
Anthropic warned that if similar standards were applied broadly across the industry, it could significantly slow or halt the deployment of frontier AI systems.
The company also defended its “defence-in-depth” safety approach, which combines technical safeguards, monitoring systems, and data retention policies intended to detect and mitigate jailbreak attempts over time.
It further stated in the release that it is working with regulators to restore access where possible and clarify the basis of the directive.
The suspension adds to a growing pattern of regulatory scrutiny faced by frontier AI developers, particularly in the United States, where agencies have increasingly asserted authority over dual-use technologies that may carry cybersecurity or national security risks.
Anthropic said it disagrees with the necessity of a full model recall based on what it describes as a narrow and non-universal vulnerability, but confirmed it is complying with the legal directive.
For now, Fable 5 and Mythos 5 remain offline pending further review. While debates continue over how far regulators should go in intervening in frontier AI systems, and what qualifies as a genuine security threat in an era of rapidly evolving model capabilities.
Comments
Sign in with Google to comment, reply, and like comments.
Continue with Google