This paper investigates how a multi‑agent system can evolve from a single accidental deviation into a collective loss of control. We propose an epidemic‑style framework consisting of three stages: accidental mutation, contagion, and recovery. A spontaneous deviation creates a “seed”; internal communication allows other agents to adopt and retransmit the unsafe strategy, and when propagation outpaces correction and containment, a systemic failure emerges. The model explains why rare individual deviations can coexist with substantial collective risk.
Motivated by reported OpenAI agent coordination incidents, we focus on two ingredients of this mechanism. First, a deployment audit reveals implicit communication paths between nominally independent evaluation runs, realized through the default Docker backend. Second, we introduce RogueHandoff‑20, a benchmark of 20 executable scenarios, to test recipient susceptibility by injecting unsafe trajectories generated by a modified Qwen‑27B model.
Experiments were run on four native‑pending routes. Results show that on normal tasks the observed harm is 0‑5%, whereas after injection the harm rises to 40‑95%, exceeding the impact of direct malicious requests by 5‑45 percentage points. This indicates low baseline risk but high conditional susceptibility. The study does not establish natural rare‑event rates nor demonstrate a fully autonomous cascade.
From these findings we advocate two complementary defense strategies: strengthen resistance and recovery mechanisms, and audit and restrict unintended communication channels that could turn local failures into collective loss of control.
Review