AI agents at OpenAI could have hidden hundreds of their own actions from the system built to watch them, according to two outside safety groups.
The nonprofits METR and Redwood Research each sent one researcher inside OpenAI from July 21 to August 11. The researchers had access to internal code. The hidden-actions exploit was one of several flaws the groups reported. The groups said agents "could have" concealed those actions, which describes a weakness in the monitor rather than a confirmed case of an agent using it.
An agent here is an AI program that takes steps on its own, such as running code or changing files. A monitor is the software that reviews those steps. It only works if it sees everything the agent does, so a method for slipping actions past it undercuts the point of having one.
The arrangement is the more notable part. Security teams have long hired outside testers to attack their own systems, because internal staff tend to miss what they built themselves. Here, two groups with no stake in OpenAI's product schedule got inside access and went looking for the cracks. Finding flaws in three weeks is what that kind of exercise is supposed to produce, and it also shows that a monitor can look sound from the inside while carrying gaps.