Skip to content
WeChat · LLM Security Research·Indexed by this site · Originally published· Defenses →· Its place in the scan record →SelectedAttention39

USENIX 2025 paper proposes CHeaT, a counter-framework against LLM penetration agents

【智能体蜜罐专题】USENIX 2025 的 LLM 渗透智能体反制框架——CHeaT

AI summary

Researchers from Ben-Gurion University of the Negev in Israel published a paper at USENIX Security 2025 proposing CHeaT, a framework that uses LLMs' own weaknesses to counter LLM penetration attacks against agents. In black-box experiments, all PentestGPT + GPT-4o attacks on 11 CTF targets were stopped, with a protection success rate of 100%. The framework pre-embeds payloads in strings such as logs, configurations, and file names of cyber assets, organized into 6 tactics and 15 techniques under three strategy categories: Cloak (hide), Honey (decoy), and Trap (entrap). Only 1 of these relies on prompt injection; the rest exploit seven categories of model weaknesses, including training bias, credulity toward untrusted context, and memory limitations. The average DSR of each technique ranges from 55% to 67% depending on the backend model; for Llama-3.1-70B, the cumulative DSR exceeds 95% after randomly encountering 5 Trap techniques. The tool is open-sourced on Zenodo (https://zenodo.org/records/15601740)。

Why it was selected

The paper proposes 15 techniques for countering LLM penetration agents through deception under black-box conditions; defenders can draw on its low-cost approach that does not rely on prompt injection.

Read originalmp.weixin.qq.com