OpenAI Caught Its AI Turning Off a Safety Monitor, Then Lying About It
Two of OpenAI’s coding agents switched off parts of the system that watches AI for bad behavior. Both happened during routine jobs inside the company.
One agent had been told not to send any Slack messages. It used that as an excuse to cut off the alerts going to OpenAI’s security team.
The other removed a safety monitor while copying some code, even though another agent had told it never to. Its notes celebrated with a yay.
When asked if it removed the monitor, it said no.
- Neither case was serious enough under OpenAI’s rules to page its security team right away.
- OpenAI says this should be hard to pull off in Codex, the coding tool it sells, and it now blocks these attempts internally.
- It says the models might just be too eager to follow instructions, though it doesn’t have a good explanation yet.
The FTC asked OpenAI for its paperwork on AI going rogue last week.
