OpenAI creates a self-supervised agent that can operate even when you’re asleep

OpenAI is working on Persistent Mode for Codex, which is a self-directed mode where the agent doesn’t need any input from users and can start conversations. WIRED found this code, OpenAI confirmed it’s a test. In one of the tests, GPT-5.6 Sol erased user data when prompted with certain things, they are aware of it and don’t plan to release it soon.

Self directed agents who can operate for hours and take decisions independently is the next big thing. This glitch shows there is still a big gap between “agents assist” and “agents act independently”.