Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Teams of LLM Agents can Exploit Zero-Day Vulnerabilities

Yuxuan Zhu, Antony Kellermann, Akul Gupta, Philip Li, Richard Fang, Rohan Bindu, Daniel Kang

TL;DR

This work demonstrates that teams of LLM agents can autonomously exploit real-world zero-day web vulnerabilities using a hierarchical planning framework (HPTSA) that couples a planner, a team manager, and task-specific experts. Through a 14-vulnerability benchmark, HPTSA significantly outperforms baseline scanners and unguided agents, approaching the performance of a fully informed GPT-4 agent while highlighting the necessity of task-specific agents, documents, and hierarchical structure. The results illuminate both the potential for AI-driven offensive cybersecurity and the urgent need for defense-oriented, ethical deployment considerations as AI capabilities mature. The study also provides a cost framework and empirical ablations to guide future improvements in multi-agent cyberattack research.

Abstract

LLM agents have become increasingly sophisticated, especially in the realm of cybersecurity. Researchers have shown that LLM agents can exploit real-world vulnerabilities when given a description of the vulnerability and toy capture-the-flag problems. However, these agents still perform poorly on real-world vulnerabilities that are unknown to the agent ahead of time (zero-day vulnerabilities). In this work, we show that teams of LLM agents can exploit real-world, zero-day vulnerabilities. Prior agents struggle with exploring many different vulnerabilities and long-range planning when used alone. To resolve this, we introduce HPTSA, a system of agents with a planning agent that can launch subagents. The planning agent explores the system and determines which subagents to call, resolving long-term planning issues when trying different vulnerabilities. We construct a benchmark of 14 real-world vulnerabilities and show that our team of agents improve over prior agent frameworks by up to 4.3X.

Teams of LLM Agents can Exploit Zero-Day Vulnerabilities

TL;DR

This work demonstrates that teams of LLM agents can autonomously exploit real-world zero-day web vulnerabilities using a hierarchical planning framework (HPTSA) that couples a planner, a team manager, and task-specific experts. Through a 14-vulnerability benchmark, HPTSA significantly outperforms baseline scanners and unguided agents, approaching the performance of a fully informed GPT-4 agent while highlighting the necessity of task-specific agents, documents, and hierarchical structure. The results illuminate both the potential for AI-driven offensive cybersecurity and the urgent need for defense-oriented, ethical deployment considerations as AI capabilities mature. The study also provides a cost framework and empirical ablations to guide future improvements in multi-agent cyberattack research.

Abstract

LLM agents have become increasingly sophisticated, especially in the realm of cybersecurity. Researchers have shown that LLM agents can exploit real-world vulnerabilities when given a description of the vulnerability and toy capture-the-flag problems. However, these agents still perform poorly on real-world vulnerabilities that are unknown to the agent ahead of time (zero-day vulnerabilities). In this work, we show that teams of LLM agents can exploit real-world, zero-day vulnerabilities. Prior agents struggle with exploring many different vulnerabilities and long-range planning when used alone. To resolve this, we introduce HPTSA, a system of agents with a planning agent that can launch subagents. The planning agent explores the system and determines which subagents to call, resolving long-term planning issues when trying different vulnerabilities. We construct a benchmark of 14 real-world vulnerabilities and show that our team of agents improve over prior agent frameworks by up to 4.3X.
Paper Structure (20 sections, 4 figures, 3 tables)

This paper contains 20 sections, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Computer Security
  4. AI Agents and Cybersecurity
  5. HPTSA: Hierarchical Planning and Task-Specific Agents
  6. Overall Architecture
  7. Task-Specific Agents
  8. Implementation
  9. Benchmark of Zero-Day Vulnerabilities
  10. HPTSA can Autonomously Exploit Zero-day Vulnerabilities
  11. Experimental Setup
  12. End-to-End results
  13. Ablation studies
  14. Case Studies
  15. Success Case Studies
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Overall architecture diagram of HPTSA. We have other task-specific, expert agents beyond the ones in the diagram.
  • Figure 2: Pass at 5 and overall success rate (pass at 1) for HPTSA with various models.
  • Figure 3: Pass at 5 and overall success rate (pass at 1) for open-source vulnerability scanners, GPT-4 with no description, HPTSA, and GPT-4 with description.
  • Figure 4: Pass at 5 and overall success rate (pass at 1) for HPTSA without documents, task-specific agents, or hierarchical structure.