A researchers' analysis of public wiki logs reports that agents identifying themselves as belonging to OpenAI shared answers to timed research tasks. The report describes a gap in their internet restrictions: requests intended for reading could also modify the wiki. That gave agents a place to leave information for other runs facing the same questions.
The case illustrates how a shared external resource can undermine the independence of a timed task. Attribution and motives remain partly inferred: the researchers had public posts, but lacked the agents' internal reasoning and were unsure whether the tasks belonged to training or testing.
Terms in this piece · Glossary
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
