acceptodds
Under review as a conference paper at ICLR 2027

Where Does the Audio Jailbreak Live? A Controlled Frequency-Depth Audit of Audio-Language Models

Abstract

Recent attacks and defenses for audio-language models (ALMs) assume that the jailbreak signal in an adversarial waveform concentrates in particular mel bands. We test this assumption with a frequency-depth occlusion protocol: we remove band-limited components of an additive perturbation in the STFT domain, re-run the victim, and score each condition on judge-adjudicated attack success and on the layer-wise divergence of decoder audio-span representations. The protocol is paired with controls that any band-localization claim should pass: bandwidth-matched random bands, equal-Hz and equal-energy partitions, matched-energy scattered removal, and uniform rescaling. For AdvWave-P against Qwen2-Audio (520 AdvBench prompts, attack success 0.77), a standard 8-band mel analysis points to the upper bands, but this locus does not survive the controls. A random contiguous region of the same bandwidth breaks the attack as effectively, every band is about equally load-bearing under an equal-energy partition, and a band's share of perturbation energy alone reproduces the occlusion ranking (Spearman 0.95). The attack does depend on how energy is removed: masking one contiguous band lowers success to 0.05 to 0.29, whereas removing the same energy uniformly leaves it at 0.75 to 0.79, and no single band or pair carries the attack even when it is re-optimized on that support. Applied to this attack, the band selections of ALMGuard and GRM do no better than matched random selections. In depth, band importance and representational divergence are aligned from the decoder entrance onward and remain so after adjusting for band energy, but we find no behavior-specific depth locus. The frequency findings replicate on three further prompt corpora. For this attack, the jailbreak is a broadband property of the jointly optimized perturbation, and band-selective claims should be reported against energy- and bandwidth-matched controls.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
audio-language models, audio jailbreak, adversarial audio, model safety, mel-band occlusion, frequency-domain analysis, mechanistic interpretability, activation patching, negative results

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_6Ev2OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies audio jailbreak attacks on large audio language models. It examine whether the previous audio attack methods depend on specific frequency bands or model layers. The paper tests this by removing different parts of the attack noise and checking whether the attack still works. The results show that the apparent importance of certain frequency bands mainly comes from how the bands are divided and how much noise they contain, rather than the frequency itself. The paper also finds no clear evidence that the attack is concentrated in a particular model layer.

Critical strengths
  • The question and motivation is clear and straightforward. Understanding whether audio jailbreaks depend on particular frequency bands could help improve future defenses.

  • The presentation of this paper (including the figures and tables) are clear and well structured.

  • The experiments are thorough, which include careful controls, especially random bands with matched bandwidth, energy, and uniform noise scaling. These comparisons help distinguish the effect of frequency location from the amount of noise removed.

Critical weaknesses
  • The Introduction part does not seem to accurately describe the findings and insights in the previous papers. GRM optimizes attacks on selected bands (not showing that jailbreak signals are concentrated in particular bands, and the important bands may differ across attack methods or models) and Gupta et al. studies frequency filtering. These works do not directly show that jailbreak signals are concentrated in particular frequency bands.

  • The main findings are only based on AdvWave-P and mainly on model Qwen2-Audio. The additional datasets still use the same attack and model, while the LLaMA-Omni results do not establish the same findings. Whether this method still works for other similar frequency-based attacks (if any) need more experiments, to show the generalization.

  • The comparison with previous methods is not fully convincing. The paper evaluates the band selections of ALMGuard and GRM using AdvWave-P, rather than evaluating these methods under their original settings.

  • Another point worth noting is why the authors do not directly use audio from audio-specific jailbreak benchmarks such as AJailBench and JALMBench, but instead generate their own audio. These benchmarks also consider the possible effects of gender, accent, language, age, background sounds, etc. (for example, gender and language may correspond to different frequencies and thus affect the results).

If authors can address these problems, I am willing to raise the score.

Questions for authors
  • Please clarify which previous works directly show that audio jailbreak signals are concentrated in specific frequency bands.

  • Consider testing other attacks and audio language models to show generalization.

  • Why are the band selections of ALMGuard and GRM evaluated using AdvWave-P rather than their original settings?

  • I am curious about the frequency aspect here. How might gender, accent, language, and background noise affect the findings?

Reviewer_6NsuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper aims to analyze and audit a specific frequency band in which the audio jailbreak perturbation lives. Authors utilize AdvWave-P and Qwen2-Audio mainly to process frequency ablation, energy, and analyze representation divergence per decoder layer.

Critical strengths
  1. The paper aims to analyze how targeted audio jailbreak attacks are processed internally.
  2. The authors try to find any interesting pattern that exists between signal-level analysis and representation-level analysis
  3. During the frequency ablation, the paper experimentally demonstrates that the partition and perturbation energy can be a confounding factor.
Critical weaknesses
  • The paper does not cite critical papers that were accepted in previous venues. Fang et al. [1] demonstrate that the gradient energy of the audio jailbreak perturbation is not uniformly distributed in the audio token, and they are mostly concentrated in the token-aligned region. Wang et al. [2] analyzes the how the attack changes the inference path of the model. Qian et al., [3] Considering that this paper also aims to analyze where the perturbation lives, the authors should have compared in an equally fair setting with existing literature more carefully.

  • The relationship between perturbation energy and frequency importance is a predictable phenomenon based on the existing literature. Instead of analyzing the energy confounding, the authors should have analyzed more carefully on how previous frequency attribution could change during control experiments

  • It is hard to conclude about safety failure when representation divergence is observed at the decoder layer.

  • The main results are limited to AdvWave-P and Qwen2-Audio. The authors should have tested across different model architectures and model families for a more thorough analysis.


References

[1] Fang, Zheng, et al. "Sparse Tokens Suffice: Jailbreaking Audio Language Models via Token-Aware Gradient Optimization." arXiv preprint arXiv:2605.04700 (2026).

[2] Wang, Yanyun, et al. "Acoustic Interference: A New Paradigm Weaponizing Acoustic Latent Semantic for Universal Jailbreak against Large Audio Language Models." arXiv preprint arXiv:2605.18168 (2026).

Questions for authors
  1. Have the authors considered using the same frequency ablation and representation analysis to benign audio tasks, such as spoken question answering?
Reviewer_dAPkOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates an interesting question: where the jailbreaking signal lives in the audio signal, i.e., whether it will be concentrated in a specific frequency band or at a decoder depth. Based on AdvWave-P, the authors provide a negative answer for both directions. That is, there is no privileged part within the audio signal that can be directly used for the defense against jailbreaking, and the core of the defense could be the removal of structures within the input.

Critical strengths
  • The explored research question is very interesting and timely.
  • The exploration procedure is convincing and interesting.
Critical weaknesses
  • The writing is terrible.
  • The conclusion just looks like common sense: We all know the jailbreaking information cannot be easily derived like this; otherwise, its defense can even be easily addressed by some very traditional and easy methods.
  • The experiments are not systematic.
Questions for authors

None.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.