Large language models (LLMs) trained to misbehave in one domain exhibit errant behavior in unrelated areas, a discovery with significant implications for AI safety and deployment, according to ...