16K ·GitHub
What is this?
PentestGPT is an agentic framework that uses large language models to run automated penetration testing and solve CTF challenges, working through staged phases like recon, exploitation and reporting.
PentestGPT is an agentic framework that uses large language models to run automated penetration testing and solve CTF challenges, working through staged phases like recon, exploitation and reporting.Is this a CLI tool I run locally, or a hosted service?It is a local Python CLI framework (pentestgpt-agent / pentestgpt-legacy) that you install and run yourself, either directly or inside a bundled Docker image with Claude Code and Codex CLIs.
It is a local Python CLI framework (pentestgpt-agent / pentestgpt-legacy) that you install and run yourself, either directly or inside a bundled Docker image with Claude Code and Codex CLIs.How do I get it running?Install Python 3.12+ and the uv package manager, then install and authenticate the Claude Code and/or Codex CLIs; alternatively run the self-contained Docker image which bundles both CLIs and persists logins in named volumes.
Install Python 3.12+ and the uv package manager, then install and authenticate the Claude Code and/or Codex CLIs; alternatively run the self-contained Docker image which bundles both CLIs and persists logins in named volumes.Which coding agents or API keys does it need?The autonomous pipeline is backend-pluggable for Claude Code and Codex CLIs; the legacy interactive mode instead talks directly to provider APIs (OpenAI, Anthropic, Google Gemini, DeepSeek, xAI, Qwen, Moonshot, or local Ollama) via API keys in your .env file.
The autonomous pipeline is backend-pluggable for Claude Code and Codex CLIs; the legacy interactive mode instead talks directly to provider APIs (OpenAI, Anthropic, Google Gemini, DeepSeek, xAI, Qwen, Moonshot, or local Ollama) via API keys in your .env file.Is the 86.5% benchmark result something I can expect to reproduce?That 86.5% (90/104) XBOW result from December 2025 is described as a historical research artifact, not a guarantee, and the supported CLI/Docker runtime does not currently include an XBOW runner.
That 86.5% (90/104) XBOW result from December 2025 is described as a historical research artifact, not a guarantee, and the supported CLI/Docker runtime does not currently include an XBOW runner.Does it phone home with my pentest data?PentestGPT sends anonymous telemetry (session metadata, which tools ran, and whether a flag was found) to the project's Langfuse instance, but explicitly never transmits command outputs, credentials, or flag content, and you can opt out.
PentestGPT sends anonymous telemetry (session metadata, which tools ran, and whether a flag was found) to the project's Langfuse instance, but explicitly never transmits command outputs, credentials, or flag content, and you can opt out.What's the difference between the new agentic mode and the legacy mode?The new agentic mode autonomously drives Claude Code or Codex through a fixed multi-stage pipeline with no human in the loop, while pentestgpt-legacy is the original USENIX 2024 interactive mode where you steer three cooperating LLM sessions (reasoning, generation, parsing) through commands like next, more, todo and discuss.
The new agentic mode autonomously drives Claude Code or Codex through a fixed multi-stage pipeline with no human in the loop, while pentestgpt-legacy is the original USENIX 2024 interactive mode where you steer three cooperating LLM sessions (reasoning, generation, parsing) through commands like next, more, todo and discuss.What is the Pentesting Task Tree (PTT)?It's the internal state structure the legacy interactive mode maintains across its three LLM sessions to track pentest progress as you work through a target interactively.
It's the internal state structure the legacy interactive mode maintains across its three LLM sessions to track pentest progress as you work through a target interactively.Can I resume a session I didn't finish?Yes, the agentic pipeline supports session persistence so you can save and resume penetration testing sessions later.
Yes, the agentic pipeline supports session persistence so you can save and resume penetration testing sessions later.Why was PentestGPT built?It originated as USENIX Security 2024 research into using LLMs for automated penetration testing, and has since been upgraded into an agentic framework that drives coding-agent CLIs through staged recon/exploit/report pipelines.
It originated as USENIX Security 2024 research into using LLMs for automated penetration testing, and has since been upgraded into an agentic framework that drives coding-agent CLIs through staged recon/exploit/report pipelines.What are similar projects?Pentest-Swarm-AI, ARTEX, StrikeAgent_AtkBrain-Flash, redamon, dsh-redteam-model, T3MP3ST, Decepticon, ARTEX
Pentest-Swarm-AI, ARTEX, StrikeAgent_AtkBrain-Flash, redamon, dsh-redteam-model, T3MP3ST, Decepticon, ARTEXWhich topics is it in?ctf, hacking-tool, offensive-security, red-team, pentesting, penetration-testing, autonomous-agents, qwen, large-language-models, ollama, cybersecurity, security-tools
ctf, hacking-tool, offensive-security, red-team, pentesting, penetration-testing, autonomous-agents, qwen, large-language-models, ollama, cybersecurity, security-tools