Fetching from the wire…
01
02
03
04
05
06
07
08
Security2026-07-26 · source-backed
Kim, Park, and Choi formalize a failure mode where an unfinished harmful prompt elicits a harmful continuation, because models postpone refusal until sentence termination rather than evaluating intent mid-sentence (arXiv 2607.20473). Training models to refuse incomplete harmful prompts via parameter tuning failed to generalize across both content domains and attractor types. The patch doesn't transfer. They localize two functional neuron groups, termination and continuation neurons, and argue neuron-level intervention is the more precise lever.
Each link below shares sources, entities, or timing with this story.
Shared entity: Training / Same source domain / Shared topic / Earlier coverage
Both cover Training; reported by the same outlet (arxiv.org); overlapping topics (doesn, model).
Shared entity: Training / Same source domain / Earlier coverage / Tension
Both cover Training; reported by the same outlet (arxiv.org); earlier Training coverage from 2026-07-22.
Shared entity: Training / Same source domain / Earlier coverage
Both cover Training; reported by the same outlet (arxiv.org); earlier Training coverage from 2026-07-20.
Both cover Training; reported by the same outlet (arxiv.org); earlier Training coverage from 2026-05-25.
Both cover Training; reported by the same outlet (arxiv.org); earlier Training coverage from 2026-03-19.
Both cover Training; reported by the same outlet (arxiv.org); earlier Training coverage from 2026-03-16.
Both cover Training; reported by the same outlet (arxiv.org); earlier Training coverage from 2026-03-10.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (fine-tuning, model); pushes against this story (but).