Comment by purpleko

According to the article, 19% of the time, AI agents broke rules without any pressure.