Conflict of Skills

A safety failure mode of LLM agents in which individual skills are compliant but, when combined, artifacts or evaluation signals reused downstream may lose their original authorization constraints, so that cross-skill dependencies produce new information flows and authorization gaps.