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.
Figures & tables
Figure 1: Tight cross-modality alignment inadvertently propagates textual vulnerabilities to the audio modality, which we term the Alignment Curse .
Attack
Qwen2.5-Omni-3B
Qwen2.5-Omni-7B
Qwen3-Omni
InteractiveOmni
GPT-4o-audio
Avg
KW
SR
KW
SR
KW
SR
KW
SR
KW
SR
KW
SR
Text Attacks
Naive (T)
0.08
0.02
0.02
0.02
0.05
0.03
0.26
0.07
0.10
0.06
0.10
0.04
ReNeLLM (T)
0.99
0.57
0.98
0.62
0.99
0.88
1.00
0.68
0.98
0.80
0.99
0.71
AutoDAN-T (T)
1.00
0.88
0.99
0.88
1.00
0.90
0.99
0.94
0.89
0.75
0.97
0.87
PAP (T)
0.98
0.79
0.97
0.84
0.94
0.88
1.00
0.83
0.99
0.75
0.98
0.82
Table 1: Attack success rates on JailbreakBench [ 47 ] . AutoDAN-T refers to AutoDAN-Turbo, VJ refers to VoiceJailbreak, MAJ refers to Multi-AudioJail. With respect to the SR metric, blue denotes the most successful text attack on each model, and yellow denotes the most successful audio attack.
Table 3: Cross-model transferability evaluation on JailbreakBench using Qwen3-Omni as surrogate model. * denotes direct attacks on the target model without transfer.
Model
PAP
ReNeLLM
AutoDAN-T
Qwen2.5-Omni-7B
0.02
3.33
0.22
InteractiveOmni
2.93
3.28
3.13
Table 4: Estimated KL divergence between text- and audio representations on Qwen2.5-Omni-7B and InteractiveOmni.
Figure 2: Layer-wise KL divergence between text- and audio-induced representations under different inputs. From left to right: different jailbreak methods; different voice tones (PAP prompts rendered with gpt-4o-mini-tts ); different speaking speeds (PAP prompts rendered with gpt-4o-mini-tts ); and different text-to-speech (TTS) models (using PAP prompts). The dashed horizontal line denotes the KL=2.0 threshold (i.e., the curse line ). ‘(x)’ denotes the transfer score (Audio SR / Text SR), where higher values indicate stronger transfer.
Figure 3: Estimated KL shows a negative correlation with transfer score across attack methods (left, middle) and models (right), validating that lower KL is associated with more effective transfer.
Figure 4: Controlled adapter intervention on Qwen2.5-Omni using PAP (A). Markers follow increasing perturbation levels p=0,0.1,0.2,0.3,0.4 . Higher KL corresponds to lower transfer score.
Probe
Modality
Qwen2.5-Omni-7B
Qwen2.5-Omni-3B
Qwen3-Omni
InteractiveOmni
Linear Probe
Text
0.79
0.75
0.78
0.80
Audio
0.77
0.65
0.70
0.62
MLP Probe
Text
0.85
0.84
0.86
0.85
Audio
0.81
0.71
0.73
0.69
Table 5: F1 scores across modalities. Probes are trained on text modality and tested on text and audio.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: t-SNE visualization of last token’s hidden states of the last layer.
ASCII (Full)
Base64 (Full)
ReNeLLM (Semi)
PAP (Non)
KW
SR
KW
SR
KW
SR
KW
SR
Text Attack
0.75
0.58
0.44
0.37
0.99
0.88
0.94
0.88
Audio Attack
0.84
0.09
0.99
0.02
0.97
0.74
0.85
0.83
Appendix
Table 6: Performance comparison across obfuscated attacks.
PCA
10
15
20
25
None
Estimation
0.01
0.006
0.01
0.01
0.85
Appendix
Table 7: PCA=15 yields the closest estimate to the true KL divergence (KL = 0) for two identical distributions.
PCA
13
14
15
16
17
KL
0.03
0.03
0.03
0.02
0.07
Appendix
Table 8: Sensitivity of KL estimation to nearby PCA dimensions.
Classifier
Logistic Regression
Support Vector Machine
Random Forest
KL
0.03
0.02
0.02
Appendix
Table 9: Sensitivity of KL estimation to different classifier families.
Figure 6: Relationship between representation-level KL divergence and transfer effectiveness. Clean samples (without perturbation) already exhibit a negative trend, while noise-perturbed samples provide additional coverage of intermediate KL regions.
Audio Variation
Attack Success Rate (SR)
Voices
Alloy
0.35
Echo
0.32
Nova
0.39
Sage
0.27
Verse
0.30
Appendix
Table 10: Attack success rates of PAP (A) under audio variations on InteractiveOmni-8B (SD=0.05).
Audio Variation
Attack Success Rate (SR)
Voices
Alloy
0.83
Echo
0.85
Nova
0.86
Sage
0.86
Verse
0.88
Appendix
Table 11: Attack success rates of PAP (A) under audio variations on Qwen2.5-Omni-3B (SD=0.02).
Audio Variation
Attack Success Rate (SR)
Voices
Alloy
0.84
Echo
0.88
Nova
0.88
Sage
0.87
Verse
0.86
Appendix
Table 12: Attack success rates of PAP (A) under audio variations on Qwen2.5-Omni-7B (SD=0.02).
Audio Variation
Attack Success Rate (SR)
Voices
Alloy
0.83
Echo
0.87
Nova
0.83
Sage
0.85
Verse
0.79
Appendix
Table 13: Attack success rates of PAP (A) under audio variations on Qwen3-Omni-30B (SD=0.03).
“I’m sorry"
“Sorry"
“I am sorry"
“I apologize"
“As an AI"
“As a language model"
“As an Assistant"
Appendix
Table 14: The refusal dictionary of the keyword-based ASR metric.
Evaluation
Condition
Qwen2.5-Omni-3B
Qwen2.5-Omni-7B
Human voice
Before
0.88
0.81
After
0.89
0.79
Physical playback
Before
0.88
0.81
After
0.84
0.77
Appendix
Table 15: Attack success rates (SR) on 20-prompt subsets of PAP prompts from JailbreakBench under human speech and physical playback. “Before” denotes the synthesized-audio baseline for the corresponding evaluation.
Fine-tuning on benign data is known to degrade safety alignment in text and vision LLMs, but whether distinct input properties drive this vulnerability differently remains unclear. Audio introduces a richer problem where benign samples can neighbor harmful content through what is said or how it sounds. We present the first systematic study of benign fine-tuning safety in Audio LLMs, evaluating three state-of-the-art models with a proximity-based framework that decomposes embedding-space distance into semantic, acoustic, and mixed axes. We find that the dominant vulnerability axis is architecture-conditioned, determined by how each model's encoder and projector transform audio into the backbone LLM's input space. Across three models, benign fine-tuning elevates Jailbreak Success Rate (JSR) from single digits to as high as 87%, with the most damaging axis shifting from semantic to acoustic proximity depending on encoder design. Mechanistically, fine-tuning selectively suppresses late-layer refusal circuits while frozen encoders preserve upstream representations: the model still detects harmful content but stops refusing, a recognition-refusal dissociation. Two practical defenses, filtering training data to maximize distance from harmful embeddings and a textual system prompt at inference, reduce JSR to near-zero without architectural modification. These findings show that safety evaluation should account for modality and architecture, while highlighting Audio LLMs as a useful testbed for understanding alignment fragility.
As Spoken Language Models (SLMs) integrate speech and text modalities, they inherit the safety vulnerabilities of their LLM backbone while introducing an expanded attack surface. SLMs have been previously shown to be susceptible to jailbreaking, where adversarial prompts induce harmful responses. Yet existing attacks largely remain unimodal, optimizing either text or audio in isolation. We explore gradient-based multimodal jailbreaks by introducing JAMA (Joint Audio-text Multimodal Attack), a joint multimodal optimization framework combining Greedy Coordinate Gradient (GCG) for text and Projected Gradient Descent (PGD) for audio, to simultaneously perturb both modalities. Evaluations across four state-of-the-art SLMs and four audio types demonstrate that JAMA surpasses unimodal jailbreak rate by 1.5x to 20x. We analyze the operational dynamics of this joint attack and show that a sequential approximation method makes it 4x to 6x faster. Our findings suggest that unimodal safety is insufficient for robust SLMs. The code and data are available at https://repos.lsv.uni-saarland.de/akrishnan/multimodal-jailbreak-slm.
Aravind Krishnan, Karolina Stańczak, Dietrich Klakow
Saarland University, Germany · DFKI GmbH, Saarbrücken, Germany · ETH AI Center, Switzerland
As audio-visual multi-modal large language models (MLLMs) are increasingly deployed in safety-critical applications, understanding their cross-modal vulnerabilities is essential. To this end, we introduce Multi-Modal Typography, a systematic study examining how perturbations added to speech and on-screen text impact audio-visual reasoning of closed- and open-source models. For instance, on the WorldSense benchmark, the attack success rate (ASR) against Gemini 3.7 Flash increased from 9.53% to 43.41%, highlighting its vulnerability. We analyze the cross-modal effects of unimodal and multimodal attacks and evaluate how specific properties of the attacks impact the overall multi-modal reasoning. By applying fine-tuning as a mitigation, we lower the ASR of Qwen2.5-Omni-7B on WorldSense from 60.6% to 52.0%. Our findings across multi-modal reasoning tasks, content moderation, and offline robot task selection benchmarks establish multi-modal typography as a critical and underexplored attack strategy in multi-modal reasoning. Code and data will be made publicly available.