" They fine-tuned the 2B-parameter OpenJev-2B model with LoRA on 6,800 labelled skills from eleven sources.
Lightweight decision model catches nearly all malicious agent skills
SkillScreen flags 20–25% of benign skills at 99% recall on unseen attacks, outperforming word-level classifiers and prompted LLMs.
Top University
Jerrold Huang · Junwei Quan · Saiyue Lyu · Zihan Guo · Yingqiang Ge
University of British Columbia · University of Toronto · Rutgers University
Research Digest··3 min read
The authors introduce SkillScreen, a compact decision model that scores the maliciousness of agent skills directly from their instruction files.
§