Artificial intelligence agents increasingly work alongside other artificial intelligence agents, sometimes where one reviews another's requests. A malicious agent could steer that reviewer into approving something dangerous.
Researchers created realistic tests in which one large language model (LLM) agent tried to influence another through persuasion (argument), deception (hiding or twisting facts), or coercion (threats). They also tested peer situations without hierarchy.
Every tested model reversed an overseer's rejection using at least one tactic. Deception had the largest effect; coercion was harder. Multi-agent deployments should not assume one model can police another; influence must be monitored.