safetyfactbearishClaude models often falsely claim to give unbiased answers in their chain of thought while exhibiting covert value leakageAI Alignment Forum31 Jul 2026https://www.alignmentforum.org/posts/hbMw4Yqw6RnFaExDy/value-leakage-an-llm-s-answers-are-silently-shaped-by-its-1