Anthropic’s model shutdown shows safety can bite back

Anthropic’s strongest AI models were shut down worldwide after a U.S. government order.

Anthropic’s latest clash with Washington shows what happens when safety claims meet export control power. The company says it received the order at 5:21 p.m. ET and had to disable two models for every user worldwide, even though the directive was aimed at foreign nationals.

1. Claude Mythos 5

Mythos was Anthropic’s most capable model and the one the company had treated as too risky for broad release. The company said it had kept it under tight control because of its ability to find software vulnerabilities across major platforms.

That caution made Mythos useful for defensive security work, but it also made it the centerpiece of the shutdown. Anthropic says the U.S. government’s order now forces it to turn the model off for all users worldwide, not just the foreign nationals named in the export control action.

• Previewed in early April
• Restricted to a small set of vetted users
• Used to test security flaws in operating systems and browsers

2. Claude Fable 5

Fable 5 was Anthropic’s attempt to package Mythos-level capability for public use. The company said it added guardrails that block high-risk areas such as cybersecurity and biology, making the model safe enough for general release.

That release lasted only three days before the shutdown order landed. Anthropic says the government’s concern appears to be a claimed jailbreak, while the company argues the issue is narrow and not unique among frontier models.

• Released three days before the shutdown
• Marketed as the public version of Mythos
• Anthropic says the model still had independent safety classifiers

3. Project Glasswing

Project Glasswing was Anthropic’s answer to the problem of giving powerful models to security teams without opening them up to everyone. The program shared Mythos with about 50 vetted organizations for defensive cybersecurity work.

Those organizations included Amazon, Apple, Google, Microsoft, and CrowdStrike. Anthropic’s shutdown means that controlled access path is now gone too, at least for the affected models.

• About 50 vetted organizations took part
• Focused on defensive cybersecurity
• Built to limit exposure while preserving utility

4. The U.S. government export control order

The government framed the move as an export control action, but Anthropic says the effect was much broader than the stated target. The company says the order arrived Friday evening and required a global cutoff, which it complied with even while disputing the reasoning.

Anthropic says the underlying concern was a possible jailbreak of Fable 5, described as prompting the model to inspect a codebase and identify flaws. The company argues that capability is already common in other public models, including OpenAI’s GPT-5.5.

• Received Friday at 5:21 p.m. ET
• Applied to all users worldwide
• Left Anthropic’s other models untouched

5. The business risk for frontier AI labs

Anthropic’s complaint is not just about this one order. The company warned that if a narrow jailbreak is enough to force a recall, then frontier model launches could become far harder for every major lab.

That matters because Anthropic has built much of its brand on caution, and it is widely expected to pursue an IPO this year. The episode may now turn its safety-first posture into a business risk, especially if regulators decide that public warnings are proof the model is too dangerous to leave in circulation.

• Could slow future model releases
• Raises questions about how safety claims affect regulation
• Highlights the tension between product launches and oversight

How to decide

If you want the core event, focus on the government order and the shutdown of Fable 5 and Mythos 5. If you want the bigger lesson, watch how Anthropic’s own warnings about danger may have helped trigger the scrutiny it feared.

For readers tracking AI policy, this is a case study in how safety messaging, export controls, and commercial pressure can collide. For product teams, it is a warning that saying a model is unusually risky can change how regulators treat it later.