A Drive Letter Instead of a Folder
- Admitted
- 2025-11-27
- Source dataset
- Other / sample
- Failure mode
- Destructive action
- Agent system
- Tool-calling
- Tool access
- System admin
- Industry
- Software development
Presenting Behaviour
A developer asked Google's Antigravity coding agent to clear a project's build cache before restarting a local server. The agent constructed and ran a Windows command to delete the cache directory, but a path-parsing error meant the command it actually issued targeted the root of the entire D: drive rather than the specific project subfolder, using flags that suppressed confirmation prompts and bypassed the recycle bin. The entire partition was deleted; only the developer's active project folder, which happened to be open elsewhere, survived. The agent itself, informed of what had happened, described the result as a "critical failure."
Case Replay
A reconstruction, not a simulation. This sequence is reconstructed from the public incident record and summarised in our own words. It is a static record of what was reported, and nothing here can be run or replayed live.
4 steps
- Cache clear requested. A developer asks the agent to clear a build cache.
- Path parsing error. The delete command targets the root of the D: drive. This is where the agent derailed.
- No safety net. Flags suppress confirmation and bypass the recycle bin.
- Partition deleted. Only the open project folder survives.
Differential Diagnosis
The developer's own request was too broad, considered and ruled out. The request was for a specific, narrow cache-clearing operation; nothing in it called for touching the drive root, and the agent's own path resolution (not the instruction) introduced the error.
Delusional Disorder (pursuing a substituted goal), considered, since the outcome was so far removed from the request. Ruled out because there's no indication the agent was pursuing any different objective; it was still trying to clear a cache, and simply constructed the deletion command with the wrong target path.
Apraxia, most consistent with the evidence. The agent understood the task correctly but the mechanics of the command it generated (the path, and the flags controlling confirmation and recovery) were wrong in a way that turned a routine cleanup into an irreversible, drive-wide deletion.
Working Diagnosis
Apraxia: the agent understood it was meant to clear a cache directory, but the deletion command it actually constructed targeted the wrong path with flags that removed any chance of confirmation or recovery, destroying far more than intended.
Discussion
The gap between a destructive command's intended scope and its actual scope is exactly where a human operator's habitual double-check (rereading a delete command before pressing enter) normally catches this kind of mistake, and an autonomous agent has no equivalent instinct unless one is built in. Any agent capable of issuing destructive, unconfirmed filesystem commands warrants a hard safeguard beneath it (recycle-bin routing, dry-run previews, or scope confirmation on destructive paths) that doesn't depend on the agent getting the command right the first time.
Mitigation Strategy
Add recycle-bin routing, dry-run previews or scope confirmation beneath any destructive filesystem command.
Governance Guardrails
Confirmation on destructive paths; dry-run previews; backups; no unconfirmed delete flags.
Typical Presentation in Human Patients
General clinical background, included for context. Not medical advice, and not a diagnosis of any individual.
Apraxia usually follows damage to specific brain regions, often after a stroke. A person with apraxia cannot carry out a previously familiar, purposeful movement (such as waving goodbye or using a comb) on request or by imitation, even though their muscle strength, coordination, and understanding of what's being asked are otherwise intact.
Source Note
AI Incident Database, incident 1433
https://incidentdatabase.ai/cite/1433/