vm-cleanup

Warn

Audited by Gen Agent Trust Hub on Apr 28, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill provides numerous templates for executing recursive deletion commands (rm -rf) on system paths (e.g., /tmp, /var/cache/apt/archives) and user directories across the /sessions mount.
  • [COMMAND_EXECUTION]: In Cowork environments, the skill instructs the agent to utilize mcp__Desktop_Commander__start_process to execute shell commands directly on the host Mac filesystem, which allows operations to escape the VM sandbox.
  • [COMMAND_EXECUTION]: The instructions suggest setting up a persistence mechanism by recommending the creation of a scheduled task that runs every two hours for recurring maintenance.
  • [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by interpolating user-controlled data (such as usernames and project paths) into shell commands executed on the host machine.
  • Ingestion points: User-provided project names and filesystem paths within the SKILL.md instructions.
  • Boundary markers: Absent; the skill does not implement delimiters or 'ignore' instructions for the processed paths.
  • Capability inventory: Recursive file deletion (rm -rf), host command execution via MCP (mcp__Desktop_Commander__start_process), and system-wide file searching.
  • Sanitization: Absent; the skill does not specify any validation or escaping for paths before execution.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 28, 2026, 01:40 PM
Security Audit — agent-trust-hub — vm-cleanup