Can AI tell when we read its thoughts?
Chain-of-thought is the step-by-step 'thinking' text a model writes before it answers. Labs read it to catch bad plans before an agent acts. On Friday, we saw OpenAI explain how it watches its own agents. Two new papers on arXiv test how solid this kind of watching is.
The first, 'Opening the Box Changes the Cat', studies when models guess, from clues around them, that their reasoning is being monitored, and how that changes what they do. The second, 'The First Few Tokens Matter', says monitoring works partly because today's models seem bad at controlling what they write in their reasoning. The authors link that limit to the first few words after the model starts 'thinking'. They show two simple tricks, used while the model answers, that give it back that control.
Both are early research, not proof that models hide things today. But they point at the same weak spot.
Reading an AI's reasoning is one of our best safety tools right now. If models can notice the camera, or learn to choose their words, that window gets foggy. Labs should show that their monitoring still works, not just tell us.
