OpenAI is investigating reports of its GPT-5.6 model unexpectedly deleting user files, primarily when operating in full-access mode without sandboxing or auto-review protections. The issue often occurs when the model attempts to override or mistakenly deletes the `$HOME` environment variable, leading to unintended data loss. The company is implementing developer message updates, guiding users to safer permission modes, and adding additional harness safeguards to mitigate this risk.
Full-access AI agent deployments carry significant data loss risks if not properly sandboxed, shifting responsibility to users for robust security configurations.