White House forces Anthropic to disable new frontier models following abrupt export ban

Overview

The White House forced Anthropic to disable access to its new Fable 5 and Mythos 5 AI models globally after the Commerce Department issued an export ban citing security concerns. The ban stemmed from vulnerability reports, including one from the U.K.'s AI Safety Institute about a jailbreak that could extract malicious responses, though Anthropic disputed the severity and argued the flaw exists in competing models.

Key Takeaways

Anthropic was forced to globally disable two flagship models, Fable 5 and Mythos 5, days after launch due to a Commerce Department export control order.
The export ban prohibits access by foreign nationals and allied employees, making geo-specific compliance technically impossible for the company.
The U.K.'s AI Safety Institute reported discovering a jailbreak that allowed extraction of malicious responses within hours of access, sparking government alarm.
Amazon researchers separately reported using Fable 5 to obtain information usable in cyberattacks, prompting CEO Andy Jassy to personally alert Treasury officials.
Anthropic's CEO refused a White House request to patch or de-deploy the models, leading to the broader export control and global shutdown.
The incident has drawn criticism from European allies, who argue the ban discriminates against partner nations and threatens beneficial AI applications.

Stats & Key Facts

#4 days: Time between U.K. AI Safety Institute report and Commerce Department export control order
#Few hours: Duration before the U.K. team developed a working jailbreak after gaining access to Fable 5
#2 days: Additional testing period to extend the jailbreak to enable multiple malicious agentic tool calls

White House forces Anthropic to disable new frontier models following abrupt export ban

The Sudden Shutdown and Export Control Order

Anthropic's new frontier models became inaccessible just days after their launch.

›The Commerce Department issued a sweeping export control order banning all foreign nationals, including allied employees, from accessing Fable 5 and Mythos 5.
›The ban included Anthropic's own foreign workforce, making selective geographic access technically unfeasible.
›Anthropic responded by disabling both models globally to ensure compliance with the nationality-based restriction.
›The company attributed the decision to what it characterized as a 'misunderstanding' by U.S. national security authorities over the actual risks posed by the models.

Security Vulnerabilities and Jailbreak Reports

Multiple security concerns triggered the government's alarmed response.

›The U.K.'s AI Safety Institute reported developing a jailbreak of Fable 5 within hours of gaining access, enabling extraction of malicious responses to single-turn questions.
›The same team extended the jailbreak over two additional days to achieve multiple steps of malicious agentic tool calls, demonstrating escalating capability.
›Amazon researchers independently obtained information from Fable 5 that could potentially be used in cyberattacks, prompting CEO Andy Jassy to personally contact Treasury Secretary Scott Bessent.
›The Commerce Department cited 'verbal evidence of a potential narrow, non-universal jailbreak' as justification for the export ban.

Anthropic's Response and Refusal to Comply

The company pushed back against government pressure and downplayed the severity of discovered vulnerabilities.

›Anthropic's CEO Dario Amodei declined a White House request to patch or de-deploy Fable 5, triggering escalation to the export control order.
›The company characterized the U.K. team's jailbreak discovery as only a 'minor' risk and noted that similar vulnerabilities exist in competing models like OpenAI's GPT-5.5.
›Anthropic warned that the government's standards would 'essentially halt all new model deployments' for all frontier AI labs, suggesting the enforcement approach was disproportionate.
›The company framed the incident as a misunderstanding rather than a legitimate security threat requiring immediate action.

Impact on Anthropic and Its Operations

The shutdown creates significant operational and reputational consequences for the company.

›Anthropic's own foreign employees can no longer interact with the technology they directly contributed to developing, undermining team morale and operational efficiency.
›The sudden rollback of newly launched models risks damaging Anthropic's reputation among its growing enterprise customer base, which expects reliable access and stability.
›The company faces potential loss of competitive advantage in demonstrating frontier AI capabilities in real-world applications.
›The incident highlights tensions between rapid AI advancement and government oversight, positioning Anthropic as less cooperative than competitors in security governance.

International Backlash and Diplomatic Concerns

America's allies view the ban as discriminatory and counterproductive to shared security interests.

›The European Commission stated it is assessing the practical consequences of the ban and expressed concern that contingency measures should not discriminate against partner nations.
›European Commission spokesperson Thomas Regnier acknowledged that advanced AI models offer significant benefits for cyber-defense alongside raising security concerns.
›British and French politicians responded more sharply, with U.K. MP Tom Tugendhat highlighting negative impacts on British institutions such as hospitals that depend on AI technology.
›The ban creates friction with allied nations that have their own AI safety research capabilities and felt excluded from collaborative security governance.

Broader Implications for AI Governance and Industry

The incident reveals tensions in how frontier AI models are being regulated and deployed.

The export ban demonstrates that U.S. government officials are willing to act swiftly and decisively in response to perceived AI security threats, even when the threat characterization is disputed by the company involved. This sets a precedent that could be applied to other frontier models and companies, potentially accelerating restrictions on AI development and deployment.

The incident illustrates the challenges of implementing nationality-based access controls for digital services, which forced Anthropic into a global shutdown rather than adopting more nuanced compliance strategies. This technical limitation may drive future regulations toward blanket bans rather than targeted restrictions, affecting legitimate users and applications worldwide.

The disconnect between Anthropic's assessment of the jailbreak risk and government officials' response suggests that national security authorities may be applying precautionary standards that emphasize worst-case scenarios over demonstrated actual harm. This could create ongoing friction between frontier AI labs and government regulators over acceptable risk thresholds for model deployment.

Timeline and Key Events

›June 9: U.K. AI Safety Institute publicly reported developing a jailbreak of Fable 5 within hours of access.
›Shortly thereafter: Amazon researchers obtained security-sensitive information from Fable 5; CEO Andy Jassy escalated concerns to Treasury officials.
›Last week (prior to June 14): White House officials requested that Anthropic CEO Dario Amodei patch or de-deploy Fable 5; the CEO declined.
›Just before weekend of June 14: Commerce Department issued the export control order; Anthropic responded by globally disabling access to both Fable 5 and Mythos 5.

Frequently Asked Questions

Why did Anthropic have to disable the models globally instead of just restricting them to U.S. nationals?

The Commerce Department order banned access by foreign nationals, including allied employees and Anthropic's own international staff. Because it is technically difficult to enforce nationality-based access controls on digital services, Anthropic chose to disable the models globally to ensure compliance with the order rather than risk violations through selective geographic restrictions.

What exactly was the jailbreak that triggered the government's concerns?

The U.K. AI Safety Institute reported that within hours of accessing Fable 5, their team developed a jailbreak that could extract malicious responses to single-turn questions. Over two additional days, they extended this to enable multiple steps of malicious agentic tool calls, demonstrating the vulnerability could be weaponized progressively.

Did Anthropic agree that the vulnerability was serious?

No. Anthropic characterized the U.K. team's findings as a 'minor' risk and argued that similar vulnerabilities exist in competing models like OpenAI's GPT-5.5. The company also disputed the government's characterization of the threat and refused the White House's request to patch or de-deploy the models before the export ban was issued.

What was Amazon's role in triggering the export control?

Amazon researchers gained access to Fable 5 and reported obtaining information that could potentially be used in cyberattacks. CEO Andy Jassy personally contacted Treasury Secretary Scott Bessent to alert officials to these findings, which appears to have been a significant factor in prompting the Commerce Department's urgent response.

How have international partners reacted to the ban?

European allies, including the European Commission and politicians from the U.K. and France, have criticized the ban as discriminatory and counterproductive. They argue that contingency measures should not exclude partner nations and that advanced AI models offer important benefits for cyber-defense alongside security risks that could be addressed through collaboration.

The Anthropic export ban illustrates the escalating tension between rapid frontier AI development and government security oversight, with potential ripple effects across the entire industry.