The Alignment Curse:模态对齐增强使文本越狱攻击更易迁移至音频
The Alignment Curse: Modality Alignment Supercharges Audio Attacks via Text Transfer
AI 导读
论文提出并验证了对齐诅咒,表明全模态模型的文本音频对齐越强,文本攻击向音频迁移越有效。对 Qwen2.5-Omni、Qwen3-Omni 等模型的黑盒评测显示,文本迁移音频攻击效果不亚于甚至常优于原生音频攻击,且在仅有音频访问时优势明显,提示现有音频安全评估可能低估来自文本模态的风险。论文已被 NeurIPS 2026 接收。
正文 · 原文
Abstract:Recent advances in end-to-end trained omni-models have substantially improved audio capabilities by strengthening text-audio modality alignment. However, whether such alignment inadvertently facilitates the transfer of safety vulnerabilities across modalities remains underexplored. This question is critical as text-based jailbreak attacks are considerably more mature than audio-based ones; if they transfer systematically, current audio safety evaluations may underestimate risks originating from the text modality. In this paper, we introduce the Alignment Curse, a formally characterized and empirically validated principle showing that stronger modality alignment enables more effective transfer of attacks from text to audio, revealing a fundamental tension between capability and safety. Motivated by this principle, we conduct a comprehensive black-box evaluation of three attack categories on recent omni-models (e.g., Qwen2.5-Omni, Qwen3-Omni): text attacks, text-transferred audio attacks, and audio attacks. We find that text-transferred audio attacks perform comparably to, and often better than, audio-based attacks, exhibiting a clear advantage under audio-only access. This suggests that text-based vulnerabilities play a pivotal role in shaping audio safety risks. Finally, we empirically analyze the relationship between modality alignment and transfer effectiveness across attack methods and models, observing consistent support for the Alignment Curse: tighter modality alignment leads to more effective cross-modality attack transfer.
| Comments: | Accepted to NeurIPS 2026 |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Sound (cs.SD) |
| Cite as: | arXiv:2602.02557 [cs.LG] |
| (or arXiv:2602.02557v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2602.02557 arXiv-issued DOI via DataCite |
Submission history
From: Yupeng Chen [view email]
[v1]
Fri, 30 Jan 2026 14:23:50 UTC (292 KB)
[v2]
Fri, 29 May 2026 23:48:47 UTC (387 KB)
[v3]
Fri, 13 Nov 2026 14:29:02 UTC (425 KB)
来源:arXiv cs.AI · arxiv.org