Morning Edition · Monday, August 24, 2026Published at 3:20 AM EDT · New York

A Safety Evaluation Paper Argues Single-Dimension Scores Hide How Aligned Models Actually Fail

The aiXamine authors give an example of a model scoring 99.3 on safety alignment while refusing one in three benign requests, and a second paper argues refusal training leaves harmful knowledge intact underneath.

The critical failure modes of deployed large language models are cross-dimensional, argue the authors of aiXamine, a unified black-box evaluation framework posted to arXiv on 24 August. Their illustrative case is blunt: a model can score 99…