xiaobai-skills
Fail
Audited by Gen Agent Trust Hub on Jun 21, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads content from several external sources to build its starter pack. While it fetches configuration from Vercel Labs' official repository (a known service), it also downloads and installs files from several unverified third-party GitHub accounts including 'mineru98', 'JimLiu', and 'mattpocock'.
- [REMOTE_CODE_EXECUTION]: Files downloaded via 'Invoke-WebRequest' from GitHub are placed directly into the agent's active skills directories (e.g., '$env:USERPROFILE.codex\skills'). Because these files contain executable instructions and scripts for the AI agent, this process constitutes the installation and execution of unverified remote code.
- [COMMAND_EXECUTION]: The skill instructions and README explicitly direct the user and the agent to run PowerShell with the '-ExecutionPolicy Bypass' parameter. This action bypasses local security controls designed to prevent the execution of unsigned or potentially malicious scripts on the host system.
- [COMMAND_EXECUTION]: The bundled 'install-xiaobai-skills.ps1' script performs broad file system operations within the user profile, including recursive deletion and moving of directories across multiple agent platforms (Codex, Claude, and general agents folders).
Recommendations
- AI detected serious security threats
Audit Metadata