OpenAI and Apollo researchers aren't concerned that current models could carry out the most serious kinds of scheming. As with much AI safety research, testing results primarily indicate risks of ...