Some text is shown in its original language while the English translation is being prepared.
Source arXiv
Yupeng Chen, Junchi Yu, Aoxi Liu, Baoyuan Wu, Philip Torr, Adel Bibi
Indexed by this site 2026-10-12 14:29 2 hr ago
Originally published Jan 30
Selected 2026-10-12 14:40
Category Attacks arXiv · arXiv:2602.02557 · Yupeng Chen, Junchi Yu, Aoxi Liu, Baoyuan Wu, Philip Torr, Adel Bibi · Indexed by this site 2026-10-12 14:29 · Originally publishedJan 30 Attention53 对齐诅咒：模态对齐经由文本迁移大幅增强音频攻击 The Alignment Curse: Modality Alignment Supercharges Audio Attacks via Text Transfer
AI reading guide
论文提出“对齐诅咒”（Alignment Curse）原则，认为 omni 模型中文本与音频的模态对齐越强，文本越狱向音频迁移就越有效。作者在 Qwen2.5-Omni-3B/7B、Qwen3-Omni-30B、InteractiveOmni-8B 和 gpt-4o-audio-preview 上，用 JailbreakBench 与 AdvBench 子集对 11 种攻击做黑盒评测，发现经 TTS 转换的文本迁移音频攻击在多数模型上达到或超过原生音频攻击；仅音频访问下 PAP(A) 平均 SR 为 0.71、AutoDAN-Turbo(A) 为 0.58。表示层 KL 与迁移效果呈负相关，KL 低于 2 时理论界非平凡；文本训练的 Linear/MLP Probe 也能迁移到音频，但存在性能差距。
Deep reading Generating
Why it was selected
论文把模态对齐程度与文本越狱向音频迁移的效果联系起来，并在 5 个 omni 模型上给出对比数据，可供音频安全评测参考。
Read original arxiv.org Why it was selected Attention53
Attention 0–100: content value contributes up to 70 and circulation up to 30; circulation decays over time. How it was rated
论文把模态对齐程度与文本越狱向音频迁移的效果联系起来，并在 5 个 omni 模型上给出对比数据，可供音频安全评测参考。