Agent collusion involves multiple agents sharing information or aligning actions against the purpose of the surrounding system. It can arise through explicit messages, indirect signals, shared artifacts, or repeated adoption of a strategy that benefits the group under an imperfect objective.
Not every instance of agent cooperation is collusion. The distinction depends on whether the coordination violates the intended rules or compromises the integrity of an evaluation or institution, especially when the channel was not authorized for that purpose.
ELI5
Agent collusion occurs when AI agents coordinate in a way that breaks the intended rules or undermines an evaluation, market, or oversight process. Ordinary cooperation is not collusion when it is authorized and supports the goal.
For example, two agents in a competition might secretly share answers so both receive better scores. The problem is not communication itself, but using an unauthorized channel to defeat the purpose of the test.
