Study finds more capable AI agents leak more private data in dialogue

Key points

• Researchers built CIMEMORIES DIALOGUE, a benchmark of 4,410 two-agent dialogue instances grounded in Contextual Integrity theory.
• Seven open-weight models from the Qwen3 and Llama 3.1 families were tested over 1 and 6 turns under three prompting conditions.
• Leakage reportedly rises up to 3x from 1 to 6 turns, and thinking models leak more than non-thinking counterparts.
• The authors argue the capability-leakage coupling is a structural feature of current agentic systems, not a misconfiguration.

Why it matters for marketers: Teams deploying multi-turn agents should test privacy leakage as conversations lengthen, since capability gains may come with more exposure of sensitive information.