Anthropic shuts down fable, mythos models after trump admin directive

At a glance:

• Anthropic disabled its Mythos 5 and Fable 5 models on Friday night following a U.S. Commerce Department export‑control order.
• The directive targets the two newest models because of a reported jailbreak that could bypass safety classifiers for cybersecurity, chemistry and biology prompts.
• Access to Anthropic’s other models remains unchanged, but the company says the pause may last until the “national security apparatus” is hardened, possibly within weeks.

What happened

Anthropic announced on Friday evening that it had abruptly disabled access to its Mythos 5 and Fable 5 large‑language models for all customers. The decision came after the company received a directive from the U.S. Commerce Department that placed the two models under export‑control restrictions, meaning they could not be used outside the United States. In a brief post, Anthropic explained that the only way to guarantee immediate compliance was to turn the models off entirely.

The shutdown affects only the two newest models; Anthropic confirmed that its other offerings, such as Claude 3 and Claude 3.5, continue to be available without interruption. The company emphasized that the move is a temporary compliance measure while the government reviews the security implications of the reported jailbreak.

Why the government intervened

According to an Axios report, officials in the Trump administration grew concerned after learning of a jailbreak that allegedly allows users to bypass the broad classifier‑based safeguards built into Fable 5. The exploit reportedly enables prompts that solicit detailed information about cybersecurity, chemistry and biology, areas the administration deems sensitive for national‑security reasons.

The Commerce Department’s order is framed as an export‑control action, a tool traditionally used to restrict the international transfer of dual‑use technologies. In this case, the government is treating the AI models as dual‑use because the alleged jailbreak could help adversaries discover software flaws or develop harmful biological agents. A source told Axios that the “national security apparatus” is being hardened and that a resolution could arrive “in the next few weeks.”

Anthropic’s response and evidence

In its announcement, Anthropic said the government had provided only “verbal evidence of a potential narrow, non‑universal jailbreak.” The company described the exploit as one that gets Fable 5 to review a specific codebase for software vulnerabilities. Anthropic’s internal assessment found that the jailbreak has so far been used to locate only “minor” and “relatively simple” software bugs.

Anthropic also noted that publicly available models such as OpenAI’s GPT‑5.5 possess comparable capabilities for code review, implying that the risk is not unique to its own systems. Nonetheless, the company opted for a precautionary shutdown to avoid any breach of U.S. export regulations and to give regulators time to finalize their security hardening measures.

Impact on customers and the wider AI market

Customers who had begun experimenting with Mythos 5 and Fable 5 will need to revert to older Anthropic models or seek alternatives while the shutdown remains in effect. For enterprises that signed up for early‑access programs, the abrupt loss of capability may delay product roadmaps that relied on the advanced reasoning and coding assistance promised by the new models.

The incident underscores growing tension between rapid AI innovation and government attempts to control potentially dangerous capabilities. Analysts predict that similar export‑control actions could become more common as other nations watch the United States’ approach to AI risk mitigation. Companies developing cutting‑edge models may need to build compliance frameworks that can react quickly to regulatory directives, potentially slowing the pace of public releases.

What to watch next

Regulators are expected to release a more detailed set of guidelines on AI export controls in the coming weeks. Anthropic has indicated it will reinstate Mythos 5 and Fable 5 once the “hardening” process is complete, but no specific timeline has been announced. Stakeholders should monitor statements from the Commerce Department, as well as any follow‑up reporting from Axios or other investigative outlets, to gauge whether additional models will face similar restrictions.

In the meantime, the broader AI community is likely to scrutinize the alleged jailbreak technique, prompting both developers and security researchers to share mitigation strategies. The episode may also accelerate efforts to develop more transparent safety classifiers that can withstand sophisticated prompt engineering attempts.