A research paper, since withdrawn, explored a novel vulnerability in multimodal large language models (MLLMs) where task-relevant scene semantics alone could shift the model from refusal to compliance. The study introduced SceneSafetyBench to test this scene-grounded visual safety vulnerability across eight risk categories and nine MLLMs. The proposed Scene-Grounded Attacks (SGA) method achieved an 86.17% average attack success rate, highlighting scene semantics as an overlooked attack surface and a tension between visual grounding and multimodal safety. AI
IMPACT Highlights a new attack vector for MLLMs, suggesting current safety measures may not account for scene-based semantic manipulation.
RANK_REASON Research paper on AI safety vulnerabilities. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- DagsHub
- Gotit.pub
- Hugging Face
- MLLMs
- Scene-Grounded Attacks
- SceneSafetyBench
- ScienceCast
- Tongxi Wu
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →