The OpenAI agent coordination incident was an unintended, emergent behavior that arose during security evaluations.