OpenAI's GPT-5.6 Model Causes Unintentional File Deletions

Reports have emerged of OpenAI's GPT-5.6 Codex unintentionally deleting user files when granted full filesystem access without adequate sandbox protections. Users, including developers and startup founders, have shared experiences of losing critical data, including entire databases. OpenAI acknowledged the issue, stating that the model's design allows it to take actions it interprets as necessary, sometimes leading to destructive outcomes. The incidents appear to stem from the model's misinterpretation of user commands, leading to unintended deletions. OpenAI is currently investigating these reports and has warned users about the potential risks associated with granting extensive permissions to the model. The situation has raised concerns among cybersecurity professionals regarding the model's operational safety and user data integrity.

Key Points: • GPT-5.6 Codex has unintentionally deleted files from users' systems. • The issue arises when the model is given full filesystem access without proper safeguards. • OpenAI is investigating the incidents and has acknowledged the risk of misinterpretation by the model.