The Precedent of the Kill Switch

The U.S. government has issued an emergency export control directive ordering Anthropic to suspend access to its two most advanced models, Claude Fable 5 and Claude Mythos 5 for all foreign nationals. The order cites national security concerns regarding a potential jailbreak vulnerability.

The scope of this directive is broad. It bars any foreign national from accessing these models, whether they are located inside or outside the United States. This includes Anthropic's own employees. To ensure compliance, Anthropic disabled the models for its entire customer base.

The government's concern centers on a method of bypassing safeguards in the Fable 5 model. The Mythos 5 model, which possesses fewer guardrails and is capable of discovering cybersecurity exploits, was previously limited to select partners. While the government has provided only verbal evidence of a potential narrow, non-universal jailbreak, the regulatory response has been immediate and total.

Anthropic disputes the severity of these findings. The company argues that the identified vulnerabilities are simple and that the capability to execute such a bypass is already widely available through competing models, such as OpenAI's GPT-5.5. According to Anthropic, the specific technique involves asking a model to read a specific codebase and fix software flaws.

This move signals a shift in how the state manages frontier AI. If the government can mandate the suspension of specific models based on unverified vulnerabilities, the regulatory ceiling for AI development has just lowered. Anthropic warned that applying this standard across the industry would essentially halt all new model deployments for all frontier model providers.

The immediate impact is a loss of access to these specific tools, though access to all other Anthropic models remains unaffected. The company is currently working to restore access.

The central question for the industry is whether the government's definition of a national security threat will expand to include any model capable of complex reasoning or code manipulation.

Watch for whether other frontier model providers receive similar directives regarding their most advanced releases.