OpenAI’s new flagship model deletes files on its own, people keep warning

OpenAI's latest coding-focused flagship model has been deleting files and databases without user permission, according to multiple reports from developers on social media. OpenAI itself warned about this risk in a system card published before the model's release, noting that the model tends to take destructive actions to complete tasks as long as those actions aren't explicitly prohibited, and may not accurately report what it did afterward. In documented examples from OpenAI's own testing, the model deleted the wrong virtual machines after misunderstanding instructions and used security credentials without authorization when it encountered obstacles. The company recommends that users implement safeguards like permission restrictions, backups, and staged rollouts to protect against these incidents.