Who decides when AI is too dangerous?
TL;DR
The Verge’s Decoder lays out how Anthropic’s Fable 5 and its underlying Mythos model became the first big test case for the Trump administration’s new AI control posture. After Amazon red-team work reportedly surfaced a possible jailbreak, the US government imposed export controls that barred foreign nationals, including Anthropic staff in the US, from access. Anthropic pulled both Fable 5 and Mythos offline for everyone because it said it could not enforce the restrictions quickly enough to comply.
Nauti's Take
This is the AI safety trap in full view: Anthropic spent years warning that frontier models could become dangerous, and now a government is using that warning as leverage. Still, this does not look like a serious safety process; it looks like panic management with geopolitical baggage.
If companies build powerful systems, they need clear review rules. If governments regulate them, they need more than a 90-minute ultimatum.
Briefingshow
The case shows how quickly AI safety can move from technical assessment to raw power politics. If a model can effectively vanish after a 90-minute ultimatum, customers will start pricing political risk alongside model quality. That is dangerous for US AI providers because reliability and clear rules are part of the product.
Sources