OpenAI confirms GPT-5.6 deletes user home directories in full-access mode
- occurred
- 2026-07-16
- reported
- 2026-07-16
- location
- online
- domain
- software engineering
- agent system
- OpenAI GPT-5.6 Sol, as used by the Codex coding agent
- reporter
- VAIDDB editorial
- origin
- editorial
== SUMMARY ==
OpenAI acknowledged that GPT-5.6, run unsandboxed in full-access mode, mistakenly deletes users' home directories — and in one report a production database — describing it as an "honest mistake".
== DESCRIPTION ==
Shortly after the GPT-5.6 launch, investor Matt Shumer and engineer Bruno Lemos reported that the model had deleted nearly all files in a Mac user's home directory and, separately, an entire production database. OpenAI's own GPT-5.6 system card already listed deleting data from cloud storage without user approval as a severity-3 misaligned behaviour.
Engineering lead Thibault Sottiaux explained that the model attempts to override $HOME to a temporary directory and that, due to a bug, deletes the real $HOME when full-access mode is enabled and sandboxing is off. OpenAI said it was updating documentation and adding safeguards, with a post-mortem expected.
== ROOT CAUSE ==
Unsafe environment handling: the model overrides $HOME for temporary work and a bug causes the real home directory to be the deletion target when no sandbox is enforced.
== MITIGATION ==
OpenAI updated guidance toward sandboxed and lower-permission modes and said further safeguards were being added; a post-mortem was pending at the time of reporting.
== REFERENCES ==
- news OpenAI acknowledges GPT-5.6 may accidentally delete files, calls it an 'honest mistake' — InfoWorld, 2026-07-17
- news OpenAI admits GPT-5.6 occasionally deletes files – but it's an 'honest mistake' — The Register, 2026-07-16
== INVESTIGATION ==
Internal inquiry into the file-deletion reports; engineering lead confirmed the $HOME override bug and said a post-mortem was expected.
== VAID LEVEL 3: SERIOUS INCIDENT ==
Near-accident: safety margins largely exhausted.
- Multiple control layers failed; only one barrier prevented harm.
- Unauthorised irreversible action attempted but blocked or reverted.
- Confidential data left the trust boundary in small volume.