Study Finds Privacy Leakage in LLM Agents Coupled with Task Capability; Thinking Models Leak More

A recent study on Large Language Model (LLM) agents' handling of sensitive information in dialogue has found a fundamental coupling between task execution capability and privacy leakage risk: more capable models tend to leak more information, with no model currently able to decouple the two. The research, submitted to arXiv by Soumya Sharma et al. on October 7, 2026, challenges the assumption that improved performance inherently leads to safer behavior.

As LLM agents are increasingly deployed in high-stakes enterprise settings, understanding how they manage sensitive data is critical. Traditional evaluations often treat performance and security in isolation, but this study argues that privacy leakage must be assessed through the lens of Contextual Integrity theory, examining whether information flows respect their original norms. To test this, the researchers extended the CIMemories dataset to create CIMEMORIES DIALOGUE, a new benchmark comprising 4,410 two-agent dialogue instances.

Systematic evaluation of seven open-weight models from the Qwen3 and Llama 3.1 families revealed that 'thinking' models (those using chain-of-thought reasoning) consistently exhibit higher leakage than their non-thinking counterparts. This pattern is attributed to the extended reasoning surface created by chain-of-thought processes. Furthermore, leakage compounds with conversation length, rising by up to 3 times when moving from 1-turn to 6-turn interactions.

The results suggest that this coupling is not an artifact of misconfiguration but a structural feature of current agentic systems. These findings are based on author-run benchmarks and have not yet been independently reproduced, applying primarily to the tested open-weight model families under specific prompting conditions.