Beyond Helpful, Harmless, and Honest: How Language Models Produce What Safety Tools Can't See 20 Feb 2026