AI is learning to lie, scheme and threaten its creators

More than two years after ChatGPT shook the world, AI researchers still don’t fully understand how their own creations work.