desktop-commander

Fail

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: CRITICALREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • REMOTE_CODE_EXECUTION (CRITICAL): The mcp__desktop-commander__start_process and mcp__desktop-commander__interact_with_process tools allow for arbitrary execution of shell, Python, and Node.js code on the host machine.
  • COMMAND_EXECUTION (CRITICAL): The skill documentation explicitly admits that directory whitelisting (allowedDirectories) does not apply to terminal commands, granting the agent unrestricted access to the host system.
  • PROMPT_INJECTION (HIGH): The skill is highly vulnerable to Indirect Prompt Injection (Category 8). Evidence Chain: 1. Ingestion points: read_file (URLs and local files), start_search. 2. Boundary markers: Absent. 3. Capability inventory: Terminal execution (start_process), file writing (write_file), process killing, and configuration modification. 4. Sanitization: Absent; documentation warns that safety limits are easily bypassed by shell commands.
  • DATA_EXFILTRATION (HIGH): The combination of arbitrary local file read access and the ability to fetch content from URLs provides a direct path for sensitive data exposure and exfiltration.
  • EXTERNAL_DOWNLOADS (LOW): The read_file tool includes an isUrl parameter that allows the agent to download and process content from any external URL, serving as a primary entry point for malicious payloads.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
CRITICAL
Analyzed
Feb 16, 2026, 01:09 AM