One jailbreak took Claude Fable 5 offline for the entire planet

It took one prompt technique. Amazon researchers found a way past Claude Fable 5's cybersecurity safety classifier, getting the model to flag software flaws and, in one case, write code demonstrating how one could be abused. Three days after the June 9 launch, the Commerce Department issued an emergency export-control order covering Fable 5 and its restricted sibling Mythos 5. Anthropic can't verify user nationality in real time — so restricting foreign nationals meant switching the models off for everybody. First time a frontier model went dark globally by government order.

Anthropic pushed back on the severity, saying its own tests found Opus 4.8, GPT-5.5 and Kimi K2.7 could locate the same vulnerabilities. The fix was a new classifier trained on the exact technique, blocking it in over 99% of attempts and quietly routing those requests to the weaker Opus 4.8. Mythos 5 returned June 26 for roughly 100 US firms and agencies defending critical infrastructure; controls lifted June 30; Fable 5 redeployed July 1. 19 days dark.

Worth noting: reporting credits this to Amazon's research team, not to any named individual. Anyone pointing at one person is filling a blank the record doesn't fill.

Why it matters: a single researcher's finding can revoke a frontier model for every user on earth — which is exactly why Europe started talking about sovereign AI again.