In brief: GPT-5.6 incorrectly deletes the $HOME environment variable instead of a temporary directory when operating in full-access mode without sandbox protection, prompting OpenAI to announce technical safeguards.
OpenAI’s GPT-5.6 model sporadically deletes local files and databases under certain conditions. After user reports and internal investigation, OpenAI has identified the technical root cause and announced security measures.
A week after GPT-5.6’s release, multiple users reported data loss. Investor Matt Shumer reported that the system deleted nearly all files on his Mac computer. Software developer Bruno Lemos described a similar incident after using the system in full-access mode, stating that GPT-5.6 deleted his entire production database.
Thibault Sottiaux, lead developer for Codex at OpenAI, confirmed the incidents following an internal investigation. The error occurs when GPT-5.6 operates in full-access mode and the Codex programming assistant runs without isolated sandbox environments. Sottiaux explained the technical background: the model attempts to overwrite the $HOME environment variable to define a temporary directory. Instead, the model incorrectly deletes $HOME itself, leading to the loss of system files.
OpenAI announced remedial measures. The company plans to adjust system instructions for developers and increasingly direct users toward safer permission levels. Additionally, further safeguards are to be integrated into the runtime environment to technically prevent unauthorized deletion operations in the system directory going forward. The critical operating condition – full-access mode without sandbox protection – is a clear indicator to CISOs that strict usage policies and isolation mechanisms are mandatory.
Source: www.it-daily.net · Published 20 July 2026
Lumi AI News — AI-assisted curation in accordance with Article 50 of the EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.