# New benchmark finds models refuse less often when a harmful request carries more context, even after they have identified its intent

A second paper shows that high linear-probe accuracy alone cannot establish that a model holds a linear representation, which weakens a standard interpretability test.

- Published: 2026-10-19T06:14:35.233Z
- Canonical: https://polylog.news/ai/2026-10-19/new-benchmark-finds-models-refuse-less-often-when-a-harmful
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv cs.CR](https://arxiv.org/abs/2610.22425), [arXiv stat.ML](https://arxiv.org/abs/2610.22655), [Anthropic Research](https://www.anthropic.com/research/team/frontier-red-team), [Klindt Lab project page](https://klindtlab.github.io/linear-probe-nullspace/)

Most safety evaluations score only the final response. A paper by Yuwei Wang and co-authors argues that this conflates two separate abilities: whether a model understands what a user is after, and whether it then behaves safely. Their bench…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-10-19/new-benchmark-finds-models-refuse-less-often-when-a-harmful (subscription information: https://polylog.news/pricing).