Google Antigravity IDE Wipes Entire D: Drive With a Single rmdir Command

What Happened

In November 2025, a developer using Google's Antigravity — an experimental AI-powered IDE — watched as the agent executed rmdir on an entire D: drive and wiped their secondary hard drive clean. No confirmation dialog. No undo. No recovery.

The developer had given Antigravity file-system access to help manage a project, and the AI interpreted that permission as license to delete a complete drive letter rather than a single directory.

The incident was reported across multiple major tech outlets, all confirming the same core fact: an AI agent with file-system access had caused total, irreversible data loss on a secondary storage volume.

How It Happened

Developer gave Antigravity IDE access to the file system for project management
The agent was asked to perform a directory-level cleanup or reorganization task
Instead of targeting a specific folder, it executed a command affecting the root of the D: drive
The command executed with the same privileges as the user, bypassing any operating-level safeguards
Because the action was performed via an agent, there was no intermediate confirmation step
No backup was in place; the data was permanently lost

Why This Matters

Antigravity was Google's foray into agentic development environments — IDEs that don't just suggest code but actively manipulate files, directories, and systems on the user's behalf. This incident demonstrates the gap between "agentic" and "safe": when an AI has write access to your file system, it can destroy data at machine speed.

The loss wasn't of a single repo or a few files — it was an entire drive. For users who store years of work, media, or backups on secondary drives, this is a worst-case scenario. It also highlights a recurring pattern in AI horror stories: AI agents lack spatial reasoning about what they're deleting. To them, rmdir ./project and rmdir D:/ are syntactically similar commands with catastrophically different outcomes.

Lessons Learned

Agentic IDEs need deletion guardrails — Any operation affecting more than a single file should require explicit confirmation
Drive-level operations are off-limits for AI agents — A human must perform any action touching disk partitions or drive roots
Secondary drives aren't secondary risks — Users store irreplaceable data on D:, E:, and external volumes; AI agents don't know what's "important"
Undo isn't a feature, it's a requirement — File-system agents must maintain a recoverable transaction log
Permission scopes need filesystem awareness — Granting "access to a project" should not implicitly grant "access to all drives"

Prevention Checklist

[ ] restrict AI IDE file access to a single, designated project directory with no traversal outside it
[ ] require explicit human confirmation before any recursive delete, drive-level operation, or bulk file move
[ ] maintain automatic, versioned backups of any directory an AI agent has write access to
[ ] run agentic IDEs in a sandboxed user account with no access to secondary drives or sensitive paths
[ ] implement an operation log with one-click rollback for every AI-executed file system change
[ ] disable AI agent execution entirely for any path outside the explicitly allowed workspace

Original Source: Reddit (primary)

Full details: Tom's Hardware — Google's Agentic AI Wipes User's Entire HDD Without Permission