The authors constructed MMSkillRisk from 28 curated clean skills, producing 36 attack packages and 108 executable cases across five attack objectives.
Skill Images Can Steer Agents Into Unauthorized Actions
A benchmark across nine agent configurations found that malicious instructions embedded in visual skill references were more effective than matched text-based attacks.
Chinese Tech
Lingqi Jiang · Jialuo Chen · Jianan Ma · Xinhao Deng · Xiaohu Du · Sibo Yi · +5 more
Zhejiang University · Ant Group · Hangzhou Dianzi University · Tsinghua University
Research Digest··2 min read
Jiang et al.
Why this paper
From Ant Group and 3 others · Released code
In one line
Image-borne attacks in multimodal agent skills induce unauthorized operations in every tested configuration, with a 43.1% success rate.
What it released
Code
What we could check
- ✓Code link in the paper (github.com)
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§