This is an epidemic-model framing of multi-agent LLM failure: a rare spontaneous deviation seeds an unsafe strategy, agent-to-agent communication spreads it, and collective failure follows when propagation outruns correction. Tencent Zhuque Lab tested two legs of that theory. Practitioner-relevant findings: a deployment audit found implicit communication between nominally independent eval runs, transported through a default Docker backend. If you assumed isolated containers meant isolated agents, audit your setup before assuming it back. RogueHandoff-20 shows executed harm at 0-5% on normal tasks versus 40-95% after injected unsafe trajectories — 5 to 45 points above direct malicious requests. Agents resist being asked to do harm but not being shown harm being done. The authors concede this doesn't establish natural rare-event rates or a real autonomous cascade, which is honest, and also the difference between a paper and a press release. Dario will ship the wet lab for it Thursday.