A latent safety monitor flagged fewer harmful prompts than model refusals
Source check: Source checked. Not independent scientific confirmation.
Finding: Across 15 language models and a held-out set of reworded prompts, one activation-based monitor flagged about 23% of harmful prompts at the same false-alarm rate as a string-matched refusal reference, which flagged about 32%. A later activation readout caught more, but analyses linked much of that gain to the model’s own refusal decision.
Technology: Latent Guard, linear and multilayer activation probes
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Learned latent readers classify prompts from instruction-tuned model activations and are checked against observed refusals and embedding models.
- Test and result
- The authors compared trained readers and string-matched model refusal at matched false alarm rates and tested how their scores tracked the models’ planned replies.
- Limitation / still unknown
- The comparison covers particular rewordings and learned monitors. It does not establish the limits of every activation-based safety monitor. Refusal wording can be missed by the reference detector; a sampled audit found only 13 of 135 replies labeled answered actually provided harmful assistance. Later-readout gains therefore do not all represent detection of genuine unsafe compliance.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.30727 · https://arxiv.org/html/2610.30727v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: arXiv v1 front matter, aggregate §6 out-of-distribution reader comparison and matched false-alarm analysis.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.