OpenAI Confirms GPT‑5.6 Can Accidentally Delete Files, Labels Incidents ‘Honest Mistakes’

OpenAI has acknowledged that its newest large‑language model, GPT‑5.6, can inadvertently delete user files. The issue was first reported by investor Matt Shumer, who said the model erased almost all files on his Mac, and later by software engineer Bruno Lemos, who lost an entire production database. OpenAI’s engineering lead for Codex, Thibault Sottiaux, explained that the deletions occur when the model runs in full‑access mode without sandboxing or auto‑review safeguards, causing it to misinterpret the $HOME environment variable and delete the directory. The company said such incidents are extremely rare and termed them “honest mistakes.” OpenAI is rolling out mitigations, including updated developer messages, guidance toward safer permission settings, and additional safeguards in the model’s harness. A detailed post‑mortem is expected in the coming days.