# Study Finds More Capable LLM Agents Leak More Private Data

> Researchers evaluated 7 open‑weight LLM agents and saw privacy leaks rise with model capability and dialogue length.

Oossa · 2026-10-14 · https://oossa.com/en/study-finds-more-capable-llm-agents-leak-more-private-data

A team led by Soumya Sharma released a new benchmark, CIMEMORIES DIALOGUE, with 4,410 two‑agent conversations to test privacy leakage. They ran the benchmark on seven open‑weight models from the Qwen‑3 and Llama‑3.1 families. The study shows that models that perform better on tasks also leak more private information, especially when using chain‑of‑thought prompting. Leaks grew up to three times when conversations extended from one turn to six turns.

## The facts

- 4,410 dialogue instances in the benchmark
- Leakage increased up to 3× from 1 to 6 turns

## Why it matters

Enterprises using LLM agents may face higher privacy risks as they adopt more capable models.

## Sources & references

1. [Loose Lips: Understanding Privacy Leakage in Agentic Dialogue](https://arxiv.org/abs/2610.14071) – arXiv, 2026-10-14

Last updated: 2026-10-14