Pentest AI Agents 🤖 Agent Open source

What it does

pentest-ai-agents turns Claude Code into an offensive-security research assistant for authorized penetration testing. It installs around fifty subagents, each carrying deep domain knowledge in one area (recon and OSINT, web and API testing, Active Directory, cloud, containers, mobile, wireless, social engineering, credential attacks, LLM red teaming, plus defensive analysis, forensics, and reporting). You describe a task and Claude routes it to the right specialist, which produces methodology, prioritised targets, and concrete next commands, with MITRE ATT&CK mappings where relevant.

What is inside

Agents are split into Tier 1 (advisory: you paste tool output and run the tools yourself) and Tier 2 (execution-capable: they can compose and run commands, but only after you declare an authorized scope, which every target is validated against, and Claude shows each command for approval). A shared scope-guard file enforces a hard-refusal list covering denial of service, mass scanning, unattended worms, false-flag operations, and safety-of-life systems, and CI checks that every Bash-capable agent carries the scope block. Supporting scripts include a persistent SQLite findings database, a session handoff report, a tool-availability doctor, and an optional installer for the underlying CLI tools. A companion MCP server project exists separately for automated pipelines.

Works with

Claude Code (Pro or Max). The agents are plain Markdown system prompts, so they can also run against any Anthropic-compatible endpoint or be copied into another runner.

How to install

curl -fsSL https://raw.githubusercontent.com/0xSteph/pentest-ai-agents/main/install.sh | bash

Maintenance and safety

MIT licensed and actively maintained, with a SHA-pinned CI validator. The README is explicit that the toolkit is for authorized testing only and requires signed rules of engagement and a defined scope; a disclaimer spells out the terms. Tier 2 agents run real security tooling, so authorization and scope discipline matter.

Who should use it

Penetration testers and red teamers with written authorization who want a structured specialist bench inside Claude Code.