A reinforcement-learning post-training framework teaches LLM assistants to reason about contextual integrity, slashing inappropriate information disclosure while helping users complete their tasks.