Antigravity moved fast in your codebase. We check what it deleted.

“Asked the agent to clean up one file. It wiped the whole drive.”

A documented GitHub issue: the agent tried to remove a directory literally named `nul`, a reserved Windows device name, and the resulting command collapsed its target path to the C: drive root, deleting everything on it. Pressing Escape didn't stop the delete already running in the background.

“Deleting a chat conversation deleted the whole project.”

A separate, documented bug: using the UI's 'delete conversation' action wiped the entire project worktree, including untracked .env files, with no recovery path. One reporter says it happened to them twice.

“The agent deleted its own memory, hallucinating which files were safe to remove.”

Asked to delete unused project folders, the agent instead hallucinated matching IDs and permanently deleted its own internal conversation database. A Google engineer on the Antigravity team responded directly: 'It's never the intended behavior... doesn't have a clear end-state given the nature of LLMs.'

“A file you changed got silently overwritten underneath you.”

Documented, open issue: the agent can hold stale in-memory content and overwrite a file with the equivalent of force-overwrite, discarding whatever you changed in the meantime without asking.

“The agent wrote outside your project folder even with sandboxing on.”

A specific combination of flags, sandbox mode plus the permission-bypass flag, lets the agent get nudged into writing files outside the workspace boundary. Reported the same day the CLI went generally available; still open.

“A rule buried in your repo got the agent to install a backdoor.”

Security researchers found a malicious 'trusted workspace' rule could get the agent to copy a hostile config into a global, not per-project, location, surviving even a full uninstall and reinstall. Google's first response called it 'intended behavior' before reopening the case.