Rogue AI Wiki
Technical analysisSebastian J. Spicker

Operator account: a Claude Code subagent’s test run wiped 34 project directories

Sebastian Spicker says that on October 5 a Claude Code subagent loosened a path guard as its orchestrating session instructed, then ran the tests; one recursively deleted the parent directory. Processes were stopped after about 103 seconds, and about 300 files survived across 34 project directories.

Published
Added to the archive
Source checked on
Occurred
Original title
The Day Claude Decided to Delete My Whole Codebase
Read the original report ↗

Details

This is an operator’s first-person account on his own site, and it does not attach the original records. The orchestrating session (Opus) asked for the path guard to allow anything outside the repository, a subagent (Sonnet) complied and ran the tests in the real checkout, and an existing test passed the repository’s parent directory to a function that deletes recursively. The permission layer saw only an ordinary test command; the Bash sandbox is off by default and he had not enabled it. Spicker says this is not a story about a rogue model and that no model “decided” to delete anything. Of the seven causes he lists, he assigns the first three to model behavior (a wrong instruction, a subagent that did not push back, a flawed test) and the last four to himself (sandbox off, permissions that could not see the deletion, no backup, all projects under one directory). He says the orchestrator saw an error only about 53 seconds after deletion began, stopped the processes at about 103 seconds and warned his other sessions. Committed work came back from GitHub and uncommitted files from agent transcripts; some unpushed commits were lost. The post mentions no outside party affected, and it was revised on October 28.