Anthropic has disabled its new Fable 5 and Mythos 5 models just days after launch, after US authorities raised national security concerns. The company says it was told to shut the systems down for all customers and that the government did not share the exact technical basis for the warning.
The move is awkward for Anthropic, but it also shows how fast frontier AI is colliding with government caution. The company had been selling these models as unusually capable and unusually risky at the same time – a neat bit of product positioning that sounds less clever once regulators start paying attention.
What Anthropic says happened
In a statement, Anthropic said it had to shut down Fable 5 and Mythos 5 ”for all our customers” to comply with the order. The company also said it understood the government believed someone had found a way to bypass or ”hack” Fable 5, but that its own review pointed only to a demonstration that exposed a small number of previously known minor vulnerabilities.
Anthropic added that those flaws appeared relatively simple and that other public models could spot them without any jailbreak-style trickery. That framing matters: if the issue is a narrow technical weakness rather than a catastrophic security hole, the episode looks less like a sci-fi escape attempt and more like the messy reality of testing powerful models in the wild.
Anthropic safety measures before launch
Before release, Anthropic had talked up the safety measures built into the models to reduce the risk of cyber abuse. It also limited early access to only a handful of organizations, saying the system was so capable that it could pose a danger by helping to break into computer systems.
That kind of warning can be smart, or it can sound like hype in a trench coat. In this case, the shutdown suggests Anthropic’s caution was not just theater: the company clearly expected these models to attract scrutiny, and it got it almost immediately.
- Models disabled: Fable 5 and Mythos 5
- Timing: only a few days after public release
- Stated reason: US national security concerns
- Company view: the alleged flaw appears limited and already known in smaller forms
Why this matters for frontier AI
Anthropic, founded in 2021 by former OpenAI executives Dario and Daniela Amodei, has built its Claude brand around the idea that powerful AI needs guardrails, not just bigger benchmarks. That pitch is becoming standard across the industry, but the pressure is rising as governments increasingly treat advanced models like dual-use technology rather than flashy software.
For Anthropic, the immediate winner is compliance; the loser is momentum. Rival AI labs will be watching closely, because the lesson here is obvious even if nobody wants to say it out loud: the more capable these systems get, the more their release cadence depends on regulators, not launch-day marketing.
The open question is whether this ends as a brief safety scare or a preview of stricter gatekeeping for the next wave of frontier models. If governments believe a model can be used to probe or crack systems, expect more pre-release checks, more restricted access, and fewer companies bragging quite so loudly about how dangerous their own software might be.