AI Safety Researchers Warn: Alignment Problem Is ‘Significantly Harder Than We Thought’