dissociation

This paper presents an empirical study on the safety risks of invisible orchestration in multi-agent LLM systems, finding that invisible orchestrators increase dissociation and suppress protective behavior, and that behavior-based evaluation is insufficient to detect internal-state risks.