Who decides when AI is too dangerous?
TL;DR
Anthropic released Fable 5 as the guarded public version of the more powerful Mythos 5. Less than a week later, the US government put export controls on both models. The reported trigger was a possible jailbreak found during Amazon red-teaming. The Trump administration then allegedly gave Anthropic a 90-minute shutdown ultimatum. The order blocked foreign nationals from accessing the models, including Anthropic employees inside the US. Anthropic took Fable and Mythos offline because it could not enforce that fast.
Nauti's Take
Anthropic is paying for years of safety marketing. When a company sells a model as potentially dangerous, nervous regulators are part of the bill.
The process is the real problem: a 90-minute ultimatum and a blanket foreign-national restriction are not a serious AI safety regime. That is policy by panic button, and it makes regulation feel arbitrary just when predictability matters most.
Briefingshow
This shows how thin the line has become between AI safety, industrial policy, and raw political leverage. If a possible jailbreak can trigger export controls, companies need red teams, formal review channels, and crisis plans for government intervention. Users should assume access to top models can now depend on geopolitics and citizenship rules.
Sources