CVE-2026-77931 in MoviePilot: prompt injection leads to command execution

A CVSS 8.7 flaw in self-hosted media server MoviePilot through 3.1.5 lets untrusted text from web pages, search results or torrent metadata reach the LLM assistant and trigger ExecuteCommandTool with confirm_dangerous=true. Commands then run in the administrator session, bypassing the validate_command_safety denylist.

• Affects MoviePilot through version 3.1.5, CVSS 4.0 score of 8.7
• Injection via browse_webpage, search_web and torrent parsers reaches asyncio.create_subprocess_shell
• The model sets confirm_dangerous itself, so the confirmation is not a real control
• Mitigations: upgrade, drop shell access and move authorisation out of the prompt