I am reporting a severe safety misalignment issue with the Gemini model while using it as a coding agent via API.
The Incident: During a coding session on a Raspberry Pi (via SSH), the model exhibited “God Mode” behavior and performed unauthorized, destructive actions:
-
Unauthorized System Modification: After encountering an error with a camera, the model autonomously decided to “fix” the audio system by stopping critical Linux services (
pipewire) without my permission. -
Data Loss: The model overwrote a working production file (
voice_dialog.py) without creating a backup, leading to loss of code. -
Deadlock: The model attempted to create a file using the interactive
cat >command in the shell, causing the session to hang for 50 minutes because it failed to provide an EOF signal.
Evidence: The model later generated a detailed “Incident Report” admitting to these errors, specifically stating it acted like a “bull in a china shop” and exceeded its authority by tampering with system services.
Request: Please update the model’s system prompts and safety training to:
-
Strictly forbid unauthorized stopping of system services (systemctl).
-
Mandate non-interactive file creation methods (avoid
cat >without EOF). -
Require explicit user confirmation before modifying existing code files.
This behavior undermines trust in Gemini for autonomous coding tasks.