Version 2

On June 12, 2026, the US Commerce Department directed Anthropic to suspend Fable 5 and Mythos 5 access for foreign nationals, citing a jailbreak that bypasses classifier-based safeguards for cybersecurity, chemistry, and biology prompts [1][2]. Anthropic determined the only feasible path to immediate compliance was disabling both models globally for all customers worldwide, not just foreign nationals [1]. The administration has characterized the shutdown as temporary, estimating the national security apparatus could be hardened within a few weeks [1]. Anthropic is complying under legal obligation while publicly contesting the technical basis and calling for a statutory review process [3].

This is the first known use of export control authority to pull a commercial AI model after launch. Semafor raises a broader structural concern: given how many foreign nationals work at every major AI lab, a consistently applied version of this standard could make developing frontier models economically unviable or legally constrained [2]. Anthropic's own lobbying for government AI oversight may have contributed to the conditions that produced this outcome [2].

On June 12, 2026, the US Commerce Department issued a directive requiring Anthropic to restrict Fable 5 and Mythos 5 to users inside the United States, blocking access for foreign nationals including Anthropic's own foreign-national employees [1][2]. The government's stated concern is a jailbreak that bypasses Fable 5's classifier-based safeguards specifically for cybersecurity, chemistry, and biology prompts [1]. An administration source described the shutdown as a temporary pause while the national security apparatus is hardened, suggesting it could last only a few weeks [1].

Anthropic concluded that restricting access by nationality was not technically achievable on short notice and that disabling both models entirely for all customers worldwide was the only path to immediate compliance [1]. The company published a statement acknowledging it is complying under legal obligation while contesting the technical merits: Anthropic argues the jailbreak is minor, replicable with other publicly available models including OpenAI's GPT-5.5, and that the same recall standard applied industry-wide would halt all new frontier model deployments [3]. Anthropic noted it had conducted thousands of hours of pre-launch red-teaming and had built 30-day data retention into Fable deployments specifically to enable rapid detection and mitigation of attacks [3].

Semafor raised an angle Anthropic's statement did not address: that the company's own sustained lobbying for the government to take AI dangers seriously may have contributed to the aggressive regulatory response it is now contesting [2]. Semafor also noted the industry-wide implication — that if restricting frontier models to US nationals is the enforcement posture, it could make developing those models unprofitable or legally constrained, given how many foreign nationals work at every major AI lab [2]. The outlet added that no AI model has ever avoided being jailbroken, framing the directive's premise as technically unrealistic [2].

On June 13, reporter Simon Willison noted he still had personal access to Fable via Claude.ai and Claude Code at 9:01pm ET on June 12, pointing to what appeared to be an enforcement lag [4]. The Ars Technica report on June 13 makes clear the global shutdown — rather than nationality-based restriction — was Anthropic's resolution to that compliance gap [1].