Diversity Is the Exploit: How Multi-Clip Videos Destabilize MLLM Safety
中文摘要
研究揭示多片段视频利用内容多样性，可绕过多模态大模型的安全防御，暴露了模型在处理复杂视频输入时的安全漏洞。
English Summary
Research reveals multi-clip videos exploit content diversity to bypass MLLM safety filters, exposing vulnerabilities in how multimodal models process complex, varied video inputs.
Original Excerpt
Disclaimer: This isn’t a peer review or an expert evaluation. It’s my personal reading notes after breaking the paper into an evidence… Continue reading on Medium »