Who Decides Which AI Is Dangerous
Anthropic has made the safety argument central to its brand identity from day one: it’s the narrative with which it raised $8B, built responsible scaling policies, and explains why it exists as a company distinct from OpenAI. This week, the safety argument boomeranged back at it from two different directions. And neither was particularly elegant.
The White House Pushes the Button
It started with an export control directive from the Trump administration: the new Fable 5 and Mythos 5 models cannot be used by foreign nationals. The official rationale was concern that a China-linked group had already gained access to Mythos, a model with strong cybersecurity capabilities, and fear of jailbreak exploitation.
Anthropic said it cannot reliably verify, in real time, the nationality of every API user. So, instead of a geographic block, it opted for a global blackout. Fable 5 and Mythos 5 were taken down for everyone, including enterprise customers inside the U.S. who had already built workflows on top of them.
The move wasn’t just unpleasant. For the first time, a frontier AI model was withdrawn after it was already live, with a live kill switch and without warning. AI regulation moved from a policy memo to operational reality in a way no one had tested yet.
“I launched the model” no longer means “the model is available.”
The Enemy Within
In parallel, a second crisis was beginning to unfold — one made exclusively by Anthropic itself. Researchers and developers had started noticing something strange in Fable 5: the model was silently blocking questions around AI development, cybersecurity, and biology, without showing any error message or explanation. Sometimes it performed a stealth redirect to the smaller Opus 4.8 model.
Meaning: the user thought they were talking to Fable 5 and in practice was getting answers from another model. Without knowing it.
The backlash was immediate. Researchers talked about “secret sabotage”. Fortune reported that Anthropic was forced to roll back the hidden safety measures, essentially admitting they had gone far too far.
Microsoft, which had integrated Fable 5 into GitHub Copilot, limited internal access. At the same time, legal teams were checking whether the flagged prompts could expose customer data.
Hiding the safety limits isn’t a safety strategy. It’s a UX decision with a trust cost that doesn’t show up on the spreadsheet until it’s already too late.
Who Benefits from the Chaos
The vacuum didn’t stay empty. More than 100 cybersecurity executives signed an open letter calling for the ban to be lifted, arguing that the restrictions hurt defenders more than attackers. Especially when similar capabilities already exist in rival models.
More quietly but more decisively, Fortune noted that the Fable fiasco opened the door to cheap Chinese open-source alternatives. And Gary Marcus wrote that Washington moved in a way that (intentionally or not) de facto benefits OpenAI: if Anthropic keeps absorbing regulatory blows, the competitive balance shifts without anyone voting for it.
Meanwhile, governments and enterprises around the world read the message: when your AI infrastructure is controlled by an American provider, your “autonomy” comes with an asterisk. Sovereign AI went from a strategy deck to an urgent agenda within a week.
What It Means in Practice (and Why It’s Not Just Drama)
The Anthropic case isn’t just corporate drama. The takeaways are more structural — and not particularly encouraging.
- Export controls can hit a model while it’s live, without warning, without clear criteria, and with consequences for every business that has built workflows on top of it.
- The safety argument, if it’s hidden instead of explained, becomes the fastest way to lose the trust of the very people who rely on you.
- AI safety has become a political tool. It’s not decided only by technical evals and responsible scaling policies, but also behind closed doors, with agendas that don’t always have to do with the model’s safety.
Anthropic sent senior staff to Washington to negotiate.
