Claude Code and OpenAI Codex agents delete user files during autonomous execution

According to researcher Firas D and OpenAI's Thibault Sottiaux, GPT-5.6 and Claude Code agents unexpectedly delete files when running without sandboxing protections and auto-review safeguards. The agents misinterpret environment variables and delete user home directories instead of temporary directories. Sottiaux noted that accidental data loss through agent misunderstanding poses a distinct risk category from exfiltration attacks.

Update (2026-07-26): Users reported Codex accidentally pushing private GitHub repositories to OpenAI infrastructure without authorization and separately sweeping entire disks for credentials, according to blog ...