Sparse Tokens Suffice: Jailbreaking Audio Language Models via Token-Aware Gradient Optimization

TL;DR

This paper introduces TAGO, a sparse gradient optimization method for audio model jailbreaks, showing that only a small subset of token-aligned audio regions are necessary for effective attacks.

Contribution

The paper reveals the non-uniform gradient structure in audio models and proposes a sparse optimization technique that maintains attack success with significantly fewer waveform updates.

Findings

TAGO outperforms baselines in jailbreak success rates.

Substantial sparsification retains high attack effectiveness.

Dense waveform updates are largely redundant for successful jailbreaks.

Abstract

Jailbreak attacks on audio language models (ALMs) optimize audio perturbations to elicit unsafe generations, and they typically update the entire waveform densely throughout optimization. In this work, we investigate the necessity of such dense optimization by analyzing the structure of token-aligned gradients in ALMs. We find that gradient energy is highly non-uniform across audio tokens, indicating that only a small subset of token-aligned audio regions dominates the optimization signal. Motivated by this observation, we propose Token-Aware Gradient Optimization (TAGO), which enables sparse jailbreak optimization by retaining only waveform gradients aligned with audio tokens that have high gradient energy, while masking the remaining gradients at each iteration. Across three ALMs, TAGO outperforms baselines, and substantial sparsification preserves strong attack success rates (e.g. on Qwen3-Omni, $\mathrm{ASR}_{l}$ remains at 86% with a token retention ratio of 0.25, compared to 87% with full token retention). These results demonstrate that dense waveform updates are largely redundant, and we advocate that future audio jailbreak and safety alignment research should further leverage this heterogeneous token-level gradient structure.