Indirect Tipping: Attackers Can Flip an LLM Agent Population Through Stepping-Stone Conventions With a Smaller Committed Minority

arXiv·low signalXBlueskyLinkedInCopy link

Flint, Aiello and Constantino ran populations of LLM agents to map critical-mass thresholds between coordination equilibria. Indirect routes through intermediate conventions needed a smaller committed adversarial minority than a direct challenge and could reach states a direct attack could not. How resistant an equilibrium is depends on its relations with the available alternatives, not on the equilibrium alone. For multi-agent deployments, securing individual agents does not secure the population's collective behavior.