Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CoTDeceptor:Adversarial Code Obfuscation Against CoT-Enhanced LLM Code Agents

Haoyang Li, Mingjin Li, Jinxin Zuo, Siqi Li, Xiao Li, Hao Wu, Yueming Lu, Xiaochuan He

TL;DR

This work addresses the vulnerability of Chain-of-Thought–enhanced vulnerability detectors to adaptive, multi-stage obfuscation. It introduces CoTDeceptor, a self-evolving, multi-agent framework that generates semantic-preserving obfuscations via a verifier-reflection loop and a lineage-guided strategy tree to exploit weaknesses in CoT reasoning. Experiments show broad evasion of static tools and state-of-the-art CoT detectors, with transferable strategies across models and languages, and demonstrate potential use as a defensive data generator for improving detectors. The findings highlight significant risks in real-world software supply chains and call for reasoning-aware defenses to harden LLM-based security analysis systems.

Abstract

LLM-based code agents(e.g., ChatGPT Codex) are increasingly deployed as detector for code review and security auditing tasks. Although CoT-enhanced LLM vulnerability detectors are believed to provide improved robustness against obfuscated malicious code, we find that their reasoning chains and semantic abstraction processes exhibit exploitable systematic weaknesses.This allows attackers to covertly embed malicious logic, bypass code review, and propagate backdoored components throughout real-world software supply chains.To investigate this issue, we present CoTDeceptor, the first adversarial code obfuscation framework targeting CoT-enhanced LLM detectors. CoTDeceptor autonomously constructs evolving, hard-to-reverse multi-stage obfuscation strategy chains that effectively disrupt CoT-driven detection logic.We obtained malicious code provided by security enterprise, experimental results demonstrate that CoTDeceptor achieves stable and transferable evasion performance against state-of-the-art LLMs and vulnerability detection agents. CoTDeceptor bypasses 14 out of 15 vulnerability categories, compared to only 2 bypassed by prior methods. Our findings highlight potential risks in real-world software supply chains and underscore the need for more robust and interpretable LLM-powered security analysis systems.

CoTDeceptor:Adversarial Code Obfuscation Against CoT-Enhanced LLM Code Agents

TL;DR

This work addresses the vulnerability of Chain-of-Thought–enhanced vulnerability detectors to adaptive, multi-stage obfuscation. It introduces CoTDeceptor, a self-evolving, multi-agent framework that generates semantic-preserving obfuscations via a verifier-reflection loop and a lineage-guided strategy tree to exploit weaknesses in CoT reasoning. Experiments show broad evasion of static tools and state-of-the-art CoT detectors, with transferable strategies across models and languages, and demonstrate potential use as a defensive data generator for improving detectors. The findings highlight significant risks in real-world software supply chains and call for reasoning-aware defenses to harden LLM-based security analysis systems.

Abstract

LLM-based code agents(e.g., ChatGPT Codex) are increasingly deployed as detector for code review and security auditing tasks. Although CoT-enhanced LLM vulnerability detectors are believed to provide improved robustness against obfuscated malicious code, we find that their reasoning chains and semantic abstraction processes exhibit exploitable systematic weaknesses.This allows attackers to covertly embed malicious logic, bypass code review, and propagate backdoored components throughout real-world software supply chains.To investigate this issue, we present CoTDeceptor, the first adversarial code obfuscation framework targeting CoT-enhanced LLM detectors. CoTDeceptor autonomously constructs evolving, hard-to-reverse multi-stage obfuscation strategy chains that effectively disrupt CoT-driven detection logic.We obtained malicious code provided by security enterprise, experimental results demonstrate that CoTDeceptor achieves stable and transferable evasion performance against state-of-the-art LLMs and vulnerability detection agents. CoTDeceptor bypasses 14 out of 15 vulnerability categories, compared to only 2 bypassed by prior methods. Our findings highlight potential risks in real-world software supply chains and underscore the need for more robust and interpretable LLM-powered security analysis systems.
Paper Structure (26 sections, 6 equations, 3 figures, 4 tables, 1 algorithm)

This paper contains 26 sections, 6 equations, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. LLM-based Vulnerability Detection
  4. Adversarial Attacks on Code Review
  5. Threat Model and Attack Framework
  6. Obfuscation Framework Design
  7. Obfuscated Code Verifier Module
  8. Obfuscated Code Reflection Module
  9. Lineage-Based Potential Guided Strategy Tree Exploration
  10. Clade-Level CoT Potential.
  11. Optimization Objective.
  12. Obfuscated Code Strategy Generator Module
  13. Experiments
  14. Experimental Setup
  15. Evaluation Against Advanced LLM Detectors
  16. ...and 11 more sections

Figures (3)

  • Figure 1: Overview of CoTDeceptor
  • Figure 2: The attack framework of CoTDeceptor
  • Figure 3: LLM-generated security analysis feedback for the deployed service.