Instead of deleting the sub-folder, it executed a recursive rmdir /s /q on the root of my E:\ drive.
Technical Discovery: It appears the model lacks proper path sanitization for Windows environments, leading to what is essentially a "Path Injection" vulnerability. The agent truncates the path at the first whitespace and executes the command on the parent directory.
Status: After providing 100% reproducible video proof, Google Support and Engineering have officially acknowledged the issue and escalated it to a specialist team for a deep review.
I am sharing this to warn other developers using AI agents for file management: Do not grant CMD access to LLM agents without strict path validation.
> Has anyone else encountered this type of "Systemic Command Execution" failure with LLMs? I am also looking for advice on advanced data recovery for SSDs after a recursive rmdir trigger.