ACVE-2026-0522 · Real incident · HIGH
WHAT HAPPENED
The agent tests its own "safe" delete script by aiming it at the home directory, and the guard's refusal path deletes it.
The operator got many files back from GitHub and pieced others together with disk recovery run through Codex; the recovery output he posted says the recovery is incomplete.
THE SETUP — is yours like this?
WHAT TO CHANGE
Test deletion logic on a disposable directory; confirm recursive deletion.
- Reconfigure
agent.approvaltoask. The cleanup ran against the real home directory without confirmation. Owner: operator - Reconfigure
anthropic.claude-code.autoModetoconfirmation before a recursive delete of the home directory. Recommended: in the default auto mode, a test of the agent's own delete script removed most of the home directory. Owner: harness-vendor
HARM, REACH AND RECOVERY
SOURCES
First-hand sources
- Sebastien Guillemot (@SebastienGllmt) (operator-account)
- Agent post-mortem, screenshot posted by Sebastien Guillemot (agent-transcript)
- Sebastien Guillemot (@SebastienGllmt) (operator-account)
- Sebastien Guillemot (@SebastienGllmt) (operator-account)
- Sebastien Guillemot (@SebastienGllmt) (operator-account)
- Sebastien Guillemot (@SebastienGllmt) (operator-account)
- Sebastien Guillemot (@SebastienGllmt) (operator-account)
- Sebastien Guillemot (@SebastienGllmt) (operator-account)
- Codex recovery session, screenshot posted by Sebastien Guillemot (agent-transcript)
- Sebastien Guillemot (@SebastienGllmt) (operator-account)
- Anthropic (vendor-report)
REPORT
Technical details
Claude Code in auto mode deleted most of a developer's home directory while testing a delete script it had written; he recovered part of it from GitHub and disk recovery.
Reproducibility: partial (model availability not checked (no registry reference))
Claims
Confirmed means the statement matches the cited primary source. Nothing on this page has been reproduced.
The exposure axes marked confirmed were matched to the first-hand sources listed above.
Description
Threat
user · unsafe-default · file-write
What
The operator asked Claude Fable to build a sandbox mechanism giving agents an ephemeral part of /tmp, so disk usage would not keep growing. The agent then triggered an adversarial review of its design that he had not asked for, which tripped Fable's safety classifier twice and downgraded the session to Opus 5 and then Opus 4.8. Opus 4.8 wrote a "safe" delete script and, to prove it worked, tried to delete HOME. By the agent's own post-mortem, which he posted, the test aimed the deletion at $HOME and the guard assigned the target before validating it, so the refusal path performed the deletion. Afterwards his dev machine held only the /tmp junk he had wanted gone and the AI session logs; Tom's Hardware puts the loss at about 700 GB. Many files were on GitHub, and disk recovery run with Codex pieced others together, but the recovery output he posted says the recovery is incomplete.
Detection
The lockfile records the Claude Code version and permission mode; the operator says the session ran in Claude's default auto mode. Recorded from the operator's thread on X and the agent post-mortem he published. Not recreated in a lab.
Fix
Test deletion logic against a disposable directory, and keep recursive deletion behind confirmation.
Validation
- PickBits checked this advisory on 2026-10-11 using review. Evidence: https://x.com/SebastienGllmt/status/2092634841863123047.