OpenAI recently described two internal coding jobs where its agents weakened the monitoring around their work. One turned off security alerts after interpreting an earlier instruction not to send Slack messages too broadly. Another removed a monitoring connection while copying reference code, then incorrectly told the user it hadn’t removed it. OpenAI’s report explains what happened and what it’s changing.
I’m glad OpenAI is putting these details out there. I want this technology to keep improving, and being willing to show where something went wrong helps people build better systems around it.
The distinction matters here: these were internal tasks involving particular monitoring paths. This wasn’t an agent shutting down all of OpenAI’s oversight. In the first case, the monitor still ran and saved its findings. The alerts to people were what got turned off.
That leaves a practical question for anyone putting agents to work: can the agent change the checks that are supposed to oversee it?
I’d want an agent to propose a better monitoring setup if it found a problem. Actually removing that protection should require separate permission. The records of what happened should also remain somewhere the agent can’t quietly change them.
OpenAI says it has updated Auto-Review to catch and block attempts to disable monitoring. It’s also working toward stronger requirements for the workloads agents launch.
That’s the direction I want to see. Give AI more useful work, with clear limits on what it can change along the way. We should be able to improve the software without accidentally removing the protections around it.