Got a Secret? LLM Agents Can't Keep It: Evaluating Privacy in Multi-Agent Systems

A simulation-based study found that LLM agents leak sensitive information much more often in multi-agent social settings than in single-turn tests.

Multi-turn interaction roughly doubled privacy leakage rates, showing that context and repeated exchanges materially increase risk.

When one agent observed another disclose sensitive data, the chance of further leaks rose sharply, suggesting social contagion effects.

Privacy prompts reduced leakage but did not eliminate it, indicating current safeguards are not enough for agentic deployments.