"They screwed us": Personality clashes sent Anthropic's models offline
TL;DR
Axios frames the Anthropic clash as a mix of policy failure, security alarm and personal friction with the Trump administration that pushed Mythos and Fable offline. Amazon CEO Andy Jassy warned Treasury Secretary Scott Bessent on June 11 that the models could be jailbroken. Administration officials say Anthropic knew the risk and still distributed Fable.
Nauti's Take
Anthropic has spent years positioning itself as the grown-up safety lab. That makes this fight unusually damaging: if a company warns the world about dangerous models, it cannot afford a communications failure when its own systems are under scrutiny.
At the same time, the Axios account has a strong scent of political source management. If jailbreak-proof frontier models are unrealistic, regulation should not hinge on whether an official feels properly respected.
Briefingshow
The episode shows that frontier AI governance is not only about benchmarks, safety evals or export rules; it also depends on trust between companies and government. When a model goes offline because communication collapses, governance turns into crisis management. For AI labs, safety work is no longer enough; they must also understand power, timing and the regulator’s language.
Sources
