← News

AI securitySources: arXiv · 16 Nov 2026

Tighter text-audio alignment helps jailbreaks cross from text to audio, researchers report

An arXiv paper argues that capability gains in omni-models come with a safety cost: better modality alignment makes text jailbreaks transfer more effectively to audio inputs.

In plain words

Omni-models handle text, audio and more in one system, and their training ties text and audio representations closely together. The paper's authors claim this coupling lets proven text-based jailbreak attacks work against the audio side too. If correct, audio safety testing alone may miss risks imported from text.

What was announced

The authors, writing on arXiv, formalize what they call the Alignment Curse: stronger text-audio modality alignment enables more effective transfer of attacks from text to audio, which they describe as a tension between capability and safety. Because text jailbreaks are far more mature than audio attacks, they argue that current audio safety evaluations may underestimate risks originating in the text modality.

In a black-box evaluation on recent omni-models including Qwen2.5-Omni and Qwen3-Omni, the team compared three attack categories: direct text attacks, text-transferred audio attacks, and native audio attacks. They report that transferred attacks performed comparably to, and often better than, audio attacks, with a clear advantage when the attacker has audio-only access. Across methods and models, they observed consistent support for the principle that tighter alignment improves cross-modality transfer. The findings come from the authors' own experiments and have not been independently confirmed.

What it does not show

  • The evidence is the paper's abstract only; full methods, model versions and result details are not available in the supplied passages.
  • All findings, including the transfer advantage and the alignment relationship, are the authors' own empirical claims and have not been independently verified.
  • The abstract names Qwen2.5-Omni and Qwen3-Omni as examples, so the full set of tested models is not stated in the supplied evidence.

For security and IT teams

Scope
The reported evaluation covers black-box attacks on omni-models such as Qwen2.5-Omni and Qwen3-Omni, not a broader audit of deployed voice assistants.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.