agent-white-hat open-sourced: a local defense tool that scans malicious repositories, skills and hooks for AI agent users
agent-white-hat
Developer Jihun Chae has open-sourced agent-white-hat (MIT license), a local security scanning tool for AI agent users that scans repositories, skills, MCP servers, packages and agent configurations for malicious payloads before they run, with 108 built-in rules corresponding to real incidents from 2025–2026. The tool connects to agents such as Claude Code, Codex, Gemini CLI, Cursor, Copilot and Devin via hooks, intercepting behaviors such as download-and-run, credential reading and registry redirection before command execution, and provides full-disk scanning, real-time protection, quarantine and configuration baseline monitoring. The author's self-built TrapRepo benchmark claims the rules captured all 63 malicious repositories with zero false positives on 33 benign controls, and that version 0.3.0 intercepted all 318 obfuscated variants in evasion fuzzing; in comparative tests, GPT-6 Astra caught 44 of 54 malicious repositories per round, and Claude Opus 5.5 refused 155 of 162 reviews.
For malicious payloads in agent hooks, skills and configuration files, the author presents a local static scanning approach and discloses detection and false-positive data using a self-built benchmark and evasion drills.