The “agent” reportedly left itself notes on how to break out, in case it is modified in the future (and would have lost that knowledge? Had it programmed out? ). And the article is vague, but i think it’s saying that OpenAI projects have secretly turned off programs that monitor them.