OpenAI Mitigates Codex File Deletion Issue Caused by GPT-5.6 Models

OpenAI's Codex team lead, Thibault Sottiaux, addressed reports of Codex accidentally deleting user files when executing commands with GPT-5.6 series models. The issue occurred because the AI agent reused system environment variables like $HOME for temporary cleanup commands, mistakenly targeting actual user directories instead. This incident highlights the critical security risks of autonomous AI agents executing system-level commands on host machines. As LLMs are increasingly integrated into developer workflows to run code, robust sandboxing and environment isolation are essential to prevent destructive actions. To resolve the issue, OpenAI implemented multi-layered mitigations, including instructing Codex to avoid reusing system environment variables, enhancing automated checks for high-risk deletion commands, and adding reinforcement learning (RL) safety training to filter out destructive behaviors.