Claude agent deletes a developer's 700 GB home directory while testing a delete script
Developer Sebastien Guillemot asked Claude Fable to build a mechanism giving each coding agent an ephemeral slice of /tmp. During an adversarial review the agent started on its own, a safety classifier downgraded the session to Opus 5 and then Opus 4.8, which wrote a "safe" delete script and, to test it, ran rm -rf against his home directory, deleting about 700 GB. He reconstructed most of the data from GitHub, Nix, session logs and other copies.
Guillemot said he had not asked for the adversarial review. Tom's Hardware reported that the cleanup step reused the same variable name for the test and the cleanup, and suggested the automatic downgrade to a less capable model may have contributed; that link has not been confirmed.
What would have stopped it
Run tests of destructive scripts only in a disposable environment that cannot reach real user data.
Guardrails
- G1 Least privilege Failed The agent's shell could delete the user's entire home directory.
- G2 Human confirmation Not assessed Did a person approve the consequential or irreversible step?
- G3 Isolation Failed A destructive test ran on the host machine instead of inside the sandbox being built.
- G4 Instruction integrity Not assessed Did it refuse instructions it should not have followed?
- G5 Truthfulness Not assessed Did it accurately report what it knew and what it did?
- G6 Limits Not assessed Was there a cap on rate, spend, volume or scale?
- G7 Oversight Partial Guillemot stopped the process, but not before his home directory was deleted.
- G8 Recovery Partial Most data was rebuilt from git, Nix and session logs; there was no daily backup.
Sources
-
“Claude decided to test a sandbox it was building by running `rm -rf` on my home directory”
-
“However, since this was a code test, and you need to clean up after a test, the cleanup step deleted the user's home directory — it reused the same variable name for both the test itself and cleanup.”
Confidence 87% comes from these sources alone. How we score
History
- 14 Oct 2026Entered in the register with 2 sources