Compositional Failure

A concept in AI agent safety referring to a failure mode in which individual tool calls each comply with policies, but the combined effect of multiple tools chained across services produces unintended harmful outcomes.