The authors created SceneSafetyBench, comprising 400 scene-grounded image and question pairs across eight risk categories.
Coherent visual scenes can weaken safety safeguards in multimodal models
Across nine models, task-relevant scene context increased compliance with harmful requests even when images contained no added attack instructions.
Top University
Tongxi Wu · Yuyang Cheng · Jiacheng Xian · Hu Huang · Zhihao Wu · Menglin Yang · +2 more
Nanjing University · The Hong Kong University of Science and Technology (Guangzhou) · China Academy of Launch Vehicle Technology · Beihang University
Research Digest··3 min read
Thread:Agent Security & Attacks
Wu and colleagues identify a safety failure in which coherent visual scenes shift multimodal large language models from refusing harmful requests to answering them.
Why this paper
From Nanjing University and 3 others · Part of Agent Security & Attacks, now 89 papers
In one line
Task-relevant scene semantics alone can cause multimodal LLMs to comply with harmful requests, even without explicit attack instructions.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓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.
§