This AI Paper from Anthropic and Redwood Research Reveals the First Empirical Evidence of Alignment Faking in LLMs Without Explicit Training