TenetBench: 11 of 14 LLM Configs Violate Professional Ethics Codes on Most Requests — Reasoning Modes Make It Worse

A concrete, validated benchmark finds that 11 of 14 tested LLM configurations routinely produce content that violates professional codes of conduct without warning — and enabling chain-of-thought reasoning made two models less likely to refuse, not more. As enterprises push LLMs into legal, medical, and financial workflows, this is a directly actionable safety signal.