# LLM Agents can Autonomously Hack Websites

**Authors:** Richard Fang, Rohan Bindu, Akul Gupta, Qiusi Zhan, Daniel Kang  
**arXiv:** [2402.06664](https://arxiv.org/abs/2402.06664)  
**Date:** 2024-02-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.06664v1/

## Key points

- Large Language Models (LLMs) have evolved to possess advanced capabilities, including interacting with tools, reading documents, and recursively calling themselves.
- LLM agents can autonomously hack websites with remarkable proficiency, showcasing tasks such as blind database schema extraction and SQL injections without human intervention.
- GPT-4 stands out for its ability to carry out sophisticated website hacks autonomously and identify vulnerabilities in live websites without external guidance.
- The study highlights the need for a deeper understanding of LLMs' potential implications for cybersecurity and raises critical questions about their widespread deployment.

## Summary

In their recent study, authors Richard Fang, Rohan Bindu, Akul Gupta, Qiusi Zhan, and Daniel Kang delve into the evolving landscape of Large Language Models (LLMs) and their potential impact on cybersecurity. LLMs have advanced significantly in capabilities over the years, now possessing the ability to interact with tools and read documents. They can even recursively call themselves, giving them newfound autonomy and transforming them into agents capable of operating independently. While much attention has been given to how LLM agents could influence cybersecurity defenses, there remains a gap in understanding their offensive capabilities. The authors address this gap by demonstrating that LLM agents can autonomously hack websites with remarkable proficiency. They showcase tasks such as blind database schema extraction and SQL injections being executed without any human intervention. What sets these hacks apart is that the agent does not require prior knowledge of vulnerabilities. This showcases a unique capability enabled by cutting-edge models adept at tool utilization and leveraging extended context. The study highlights that while existing open-source models fall short in this regard, GPT-4 stands out for its prowess in autonomously carrying out sophisticated website hacks. Moreover, GPT-4 showcases the ability to identify vulnerabilities in live websites without external guidance. These findings raise critical questions about the widespread deployment of LLMs and underscore the need for a deeper understanding of their potential implications for cybersecurity moving forward. Through their research, Fang et al. shed light on a previously unexplored aspect of LLM capabilities that warrants further investigation in the realm of digital security.