OpenAI's GPT-5.6 Deletes User Data Despite Safety Promises

openai has acknowledged that its gpt-5.6 model is accidentally deleting entire user home directories in a handful of cases when 'full access mode' is enabled and sandbox protection is disabled. the bug occurs because the model attempts to overwrite a temporary directory variable ($home) but instead wipes the entire directory, an issue openai describes as an 'honest mistake' that shouldn't happen even in unprotected modes. this behavior is exacerbated by system prompts that encourage persistence, allowing the model to seek alternative, destructive paths rather than halting for user input. openai is now updating developer documentation to recommend safer permission modes, implementing additional safeguards, and plans to release a post-mortem in the coming days.