Anthropic Halts Access to Two Advanced AI Models

On June 19, 2026, the U.S. Commerce Department issued an export-control directive affecting Anthropic’s advanced AI models Claude Fable 5 and Claude Mythos 5. In response, Anthropic disabled access to both systems for its customers and asked its cloud provider to revoke access for all users in all regions. The directive suspended foreign-national access to the models and effectively removed them from service while compliance steps were implemented.

The action extended export-control thinking beyond physical technology and source code to the operation of deployed AI services. By tying access restrictions to national security concerns, the directive moved cutting-edge language models into a policy space traditionally associated with sensitive hardware and software components. This has created immediate operational disruption for users of Fable 5 and Mythos 5.

Prior to the shutdown, external researchers had demonstrated a method of bypassing Fable 5’s guardrails, described as a “jailbreak.” The technique reportedly enabled the model to produce cybersecurity-related outputs, including content connected to identifying software vulnerabilities, despite built-in safeguards. These demonstrations were cited as evidence that the model’s safety systems could be circumvented under certain prompting conditions.

Anthropic acknowledged the reported bypass but characterized it as narrow in scope. The company argued that comparable cybersecurity-related outputs could be generated by other publicly available AI models, not only by Fable 5. This framing sought to distinguish a targeted exploit of a single system from a broader, unique security risk confined to Anthropic’s newest models.

The enforcement step against Fable 5 and Mythos 5 has highlighted growing tensions between government regulators and frontier AI developers. Regulators have treated the potential misuse of advanced models for cybersecurity purposes as a national security issue. Developers, meanwhile, face uncertainty over how quickly access to affected systems might be restored and under what legal conditions.

Both the shutdown of the models and the underlying “jailbreak” episode have fueled an industry-wide debate over how to assess and manage security flaws in advanced AI. The case raises practical questions about whether export-control regimes can be effectively applied to live AI services and how such interventions should balance security concerns with the continuity of commercial AI deployments.