PentestGPT is an automated penetration testing agentic framework powered by large language models, published as research at USENIX Security 2024. In its current 1.0 form it works as an autonomous agent: it drives tools such as Claude Code or Codex through a multi-stage pipeline, from reconnaissance to exploitation to a walkthrough for capture-the-flag challenges, or from asset discovery to vulnerability identification to a written report for penetration tests, feeding each stage’s findings into the next. The project has more than 16,000 stars on GitHub and was updated in 2026.
Penetration testing is expert, step-by-step work: enumerate, probe, exploit, document. PentestGPT structures an LLM’s reasoning around that workflow instead of leaving it as a free-form chat. The agent plans the engagement, runs standard security tools, interprets their output and decides the next action, with a human able to supervise or let it run autonomously. It grew out of an award-winning research system that demonstrated LLMs could meaningfully assist on real CTF and pentesting tasks, and the agentic upgrade turns that research into a pipeline that produces concrete artifacts, including final reports, rather than scattered suggestions.
PentestGPT is free and open source under the MIT license. Your only costs are the LLM services the agent uses. Use it only on systems you own or are authorized to test.
PentestGPT is for security professionals, CTF players and researchers who want to see how far an LLM-driven agent can take penetration testing workflows, and to automate the repetitive early stages of authorized assessments.