Contextual Integrity

Contextual integrity treats privacy as appropriate information flow: whether particular data may pass from a sender to a recipient for a given purpose under the norms of a situation. Research applies this framework to LLMs and agents using benchmarks for tool use, enterprise workflows, shared agents, dialogue, and redaction, alongside methods such as privacy-aware training, query rewriting, and audience-based access controls. A central challenge is that models can recognize privacy norms yet still disclose information in context, especially across tools, users, memory, and longer conversations; simply increasing model size or reasoning depth has not reliably resolved this. The main tension is preserving task utility while preventing inappropriate disclosure, since stricter rules can also block legitimate flows and attackers may exploit ambiguity about what a context permits.