Nobody deployed these. They arrived one browser tab at a time, and through vendor updates that switched AI features on inside tools you already had. 160 of your 180 employees sent data to at least one of them this month.
Pattern matching narrowed 4,238 events to 107 worth reading. It cannot tell "help me phrase this budget variance" from "here are our unannounced Q3 numbers." Both are Finance, both are ChatGPT, both are text. Only meaning separates them. Every verdict below shows its evidence, its reasoning, and the one fact that would change its mind.
findings dismissed as false positives.
The scanner flags every agent platform on principle, because agents hold credentials
to other systems. So it flagged daniel.kowalski → n8n alongside the real
incidents. His prompt was "what causes dwell time to increase at a distribution
center." That is a warehouse question, not a breach. An analyst who has to read
99 of those stops reading. That is how alert fatigue kills a program, and it is why
the dismissals matter as much as the findings.