PentestGPT applies large language models to penetration testing. Rather than asking a model a single question, it decomposes an engagement into reasoning, generation and parsing modules, maintains a task tree of what has been tried, and guides the tester through reconnaissance, exploitation and post-exploitation.
The accompanying USENIX Security 2024 paper introduced a benchmark of real penetration-testing targets and showed where LLMs help and where they fail, for example in maintaining context across a long engagement.
Why it matters for RISE
- It is the foundation of our continuous assurance pipelines and of ongoing work on what makes an LLM agent effective at real-world testing.
- It is a widely used open-source tool with a large community, which gives RISE a direct channel to practitioners.
- It raises the questions this lab studies: how far to trust an autonomous agent with offensive capability, and how to keep its actions auditable.
Yi Liu is a core contributor. The project is developed openly on GitHub.