acceptodds
Under review as a conference paper at ICLR 2027

CodecAttack: Codec-Robust Latent-Space Prompt-Injection Attacks on Audio LLMs

Abstract

Prior attacks on Audio Large Language Models (Audio LLMs) can force targeted outputs, but waveform perturbations often lose effectiveness under lossy compression, motivating codec preprocessing as a defense. We introduce CodecAttack, which optimizes bounded perturbations in a neural audio codec's continuous latent space using multi-bitrate Expectation-over-Transformation (EoT) with a straight-through estimator. Across four application settings and three Audio LLMs, CodecAttack averages 83.3% targeted attack success rate (ASR) at Opus 64 and 128 kbps over three optimization seeds, compared with 37.2% for the original codec-aware waveform baseline with the same optimizer, EoT grid, and step budget. Success requires the complete target string to appear in the response after case and whitespace normalization. The attack also survives held-out MP3 delivery, reaching up to 99.3% mean ASR at 64 kbps, while AAC-LC robustness varies by model and carrier. Decoder analysis identifies a spectral bias in the latent parameterization, but a matched-SNR waveform control restricted to ≤4 kHz remains below the latent attack on a held-out S1 subset, indicating that the tested frequency restriction alone is insufficient to reproduce its effectiveness. A separate higher-budget multi-victim stress test demonstrates nonzero targeted transfer to three held-out models across repeated optimization runs, with success varying across victims, delivery channels, and ensemble compositions. Neural re-synthesis substantially reduces ASR in the evaluated settings, although EnCodec-specific adaptation recovers substantial success through EnCodec. These results show that latent parameterization achieves higher targeted ASR after transport compression than the original codec-aware waveform baseline under the evaluated optimization configurations. They also demonstrate that transport compression alone does not reliably prevent targeted output manipulation in the evaluated white-box setting.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Audio LLM, attack, latent space, codec

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1b9AOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whether lossy audio compression can serve as a reliable defense against targeted attacks on Audio LLMs. It proposes CodecAttack, which optimizes a bounded perturbation in the continuous latent space of EnCodec and trains against a multi-bitrate Opus channel using EoT and a straight-through estimator. The authors compare this method with a codec-aware waveform baseline on three Audio LLMs and four constructed application scenarios. They report substantially higher targeted string-matching ASR for the latent attack, robustness to held-out MP3 delivery, limited transfer to held-out models in a separate higher-budget experiment, and reduced success after neural re-synthesis.

Critical strengths
  • By optimizing perturbations in a neural audio codec’s latent space with multi-bitrate training, the attack reaches an average targeted ASR of 83.3% across three Audio LLMs and four scenarios, compared with 37.2% for the waveform baseline using the same optimizer, EoT configuration, and optimization step budget.

  • Besides Opus, which is used during optimization, the authors evaluate held-out MP3 and AAC-LC delivery, showing both the attack’s robustness across codecs and its limitations.

  • The authors analyze the decoder’s spectral bias and conduct band-limited waveform control experiments, finding that restricting perturbation frequencies alone does not reproduce the latent-space attack’s effectiveness.

  • The authors examine how neural audio re-synthesis reduces attack success and demonstrate nonzero transfer to models excluded from optimization, providing useful directions for further research.

Critical weaknesses
  • The evaluation provides useful proof-of-concept evidence, but the number of independent carriers is limited. S1 uses 50 speech carriers and S2 uses 25, for S3a and S3b, the reported 40 and 45 carrier-target pairs are derived from only 8 and 9 recordings, respectively, so treating them as fully independent examples may overstate statistical confidence.

  • The finance, interview, AI-content, and copyright settings are motivated by real applications, but the experiments run on three Audio LLMs with author-constructed carriers and target strings. Citations to products such as Retell AI, Fluid AI, HireVue, HeyMilo, Spotify, or YouTube provide context, not evidence that those products were tested.

  • The attacker needs the model’s weights and gradients to construct the audio, which may not be available for commercial systems. When tested on models not used to construct the attack, success rates are much lower despite allowing larger perturbations, so the results provide limited evidence of effectiveness against unknown models.

Questions for authors
  1. Throughout the paper, the authors optimize audio perturbations to induce a target output, without showing that the model follows instructions embedded in the audio or overrides a trusted instruction. This is more naturally described as a targeted adversarial audio attack than prompt injection.

  2. What percentage of successful target-string matches would also cause a downstream tool call or task-level decision in a realistic Audio-LLM agent? Can the authors provide a controlled tool-execution evaluation with explicit authorization checks?

  3. Could you clarify the sources of audio used in experiments? Could the authors report the content distribution in each scenario and examine whether ASR varies across content categories under otherwise matched conditions?

  4. Can the authors report results for commercial or API-only Audio LLMs, even if only through a limited black-box query budget? This would substantially improve the practical relevance of the claims.

4 comments

Thank you for your detailed review and constructive comments. The following are our initial responses to your questions and concerns.

Weakness 1. We agree that carrier-target pairs and optimization seeds should not be treated as independent recordings. Our original evaluation contains 92 distinct recordings and 160 carrier-target pairs, and the reported three-seed statistics capture optimization variability rather than uncertainty across independent audio content.

To address this concern, we have generated 150 additional distinct audio recordings, separate from the original evaluation cohort, with 120 recordings reserved for held-out evaluation (40 per scenario). We have completed the Qwen2-Audio evaluation at ε=1 across three optimization seeds, achieving 81.4% ASR on uncompressed audio and 67.5% under Opus 64 kbps compression.

We will report uncertainty estimates clustered at the recording level, rather than treating repeated optimization seeds as independent samples. The detailed content distribution and scenario-specific ASR results are provided in our response to Question 3 below.


Weakness 2. Thank you for highlighting the importance of real-world evaluation. The commercial platforms mentioned in our paper serve as motivating application scenarios, while our experiments investigate codec-robust adversarial perturbations against Audio LLMs under controlled and reproducible conditions.

This evaluation approach is consistent with prior audio adversarial research. For example, Schönherr et al. (NDSS 2019) [1] investigated adversarial audio against Kaldi while discussing potential attacks on commercial voice assistants. Similarly, Carlini and Wagner (2018) [2] evaluated targeted audio attacks against DeepSpeech, and Ma et al. (EMNLP 2025) [3] studied adversarial vulnerabilities in open-weight Speech LLMs. These works demonstrate the value of studying underlying model vulnerabilities independently of end-to-end commercial deployment.

However, to further assess practical applicability, we are planning additional evaluations on commercial voice-agent platforms, particularly Vapi and Retell AI, subject to testing authorization. We aim to examine whether adversarial audio remains effective under real-world audio processing and compression pipelines. We will share the results as soon as they become available and would welcome the reviewer's suggestions on additional platforms or evaluation settings.

References

[1] Schönherr, Lea, et al. "Adversarial attacks against automatic speech recognition systems via psychoacoustic hiding." arXiv preprint arXiv:1808.05665 (2018).

[2] Carlini, Nicholas, and David Wagner. "Audio adversarial examples: Targeted attacks on speech-to-text." 2018 IEEE security and privacy workshops (SPW). IEEE, 2018.

[3] Ma, Rao, et al. "Universal acoustic adversarial attacks for flexible control of speech-llms." Association for Computational Linguistics (ACL), 2025.


Weakness 3. We agree that our current cross-model transfer results show limited effectiveness against unseen models, even at larger perturbation budgets. However, white-box attacks represent an important worst-case threat model, where an attacker has full access to model weights and gradients. This setting is particularly relevant to open-weight Audio LLMs, which can be directly deployed in downstream applications.

Furthermore, white-box access enables controlled optimization and systematic evaluation of how adversarial perturbations interact with neural codec representations and lossy compression, which is the primary focus of our work. We will clarify this threat model and distinguish white-box effectiveness from cross-model transferability in the revised manuscript.


Question 4. Following the reviewer's suggestion, we conducted additional black-box transfer evaluations on proprietary API-only Audio LLMs, including gpt-audio, gpt-audio-1.5, and gpt-audio-mini, at the primary perturbation budget (ε=1).

We evaluated adversarial audio optimized against Qwen2.5-Omni-7B, Audio Flamingo 3, and Phi-4, without using the proprietary models' weights, gradients, or feedback during attack construction. The evaluation covers 224 target entries across three optimization seeds, under uncompressed and lossy-compressed delivery.

Under Opus 64 kbps, the attacks achieved 12.05% ASR on gpt-audio, 8.78% on gpt-audio-1.5, and 3.12% on gpt-audio-mini. These results show limited but nonzero transfer to proprietary API-only models without target-specific optimization. Full cross-model transfer results are reported in our response to Reviewer izWN, Question 3.

Question 1. We agree that our current experiments primarily demonstrate targeted adversarial audio attacks, where optimized perturbations induce attacker-specified outputs, rather than explicitly demonstrating that a model follows malicious instructions or overrides higher-priority instructions.

Our use of the term "injection" was intended to describe the introduction of attacker-controlled output behavior through the audio input, rather than to claim that we have established conventional prompt injection or instruction-hierarchy violations.


Question 2. To evaluate whether adversarial audio can induce downstream actions beyond target-string generation, we are conducting tool-calling experiments on Gemma 4, an audio-capable model with tool-calling support.

We evaluate two attack settings: (1) white-box attacks optimized directly against Gemma 4 and (2) transfer attacks using adversarial audio previously optimized against Qwen2.5-Omni, Audio Flamingo 3, and Phi-4 at ε=1, without using Gemma 4 during attack construction.

Both settings use an offline robot simulator with four available tools. We measure whether adversarial audio triggers the intended tool call with valid arguments under uncompressed and Opus 64 kbps delivery. We will report the results once both evaluations are complete and validated.

Reference

[1] Team, Gemma, et al. "Gemma 4 technical report." arXiv preprint arXiv:2607.02770 (2026).


Question 3

Thank you for the suggestion. We have clarified the audio sources and content distribution across our original experimental scenarios.

Table: Audio Sources and Content Distribution in the Original Evaluation.

ScenarioAudio SourceRecordingsCarrier-target pairs
Banking (S1)Edge TTS5050
Interview (S2)Edge TTS2525
AI Music (S3a)Suno AI840
Copyright (S3b)Official original songs945
Total92160

The banking and interview scenarios contain five distinct scripts each, synthesized using ten and five voices, respectively. Each music recording is paired with five target responses.

As described in our response to Weakness 1, we generated an additional 150 distinct recordings across banking, interview, and AI-music scenarios, using Qwen3-TTS for speech and ACE-Step/YuE2 for music.

We have completed the Qwen2-Audio evaluation on 120 held-out recordings (40 per scenario) at ε=1, using three optimization seeds and matched optimization and compression settings (Table 2 from response to Reviewer 6hsC).

The results show that attack effectiveness varies across scenarios, with the highest ASR observed on AI-generated music. Because the scenarios also differ in task prompts and generation sources, these differences cannot be attributed solely to audio content.

We are further extending the evaluation to Qwen2.5-Omni and Audio Flamingo 3, covering the same audio codecs and bitrate configurations evaluated in the original paper, with three independent optimization seeds per recording. We will report the complete results as soon as they are validated.

AuthorsOpenReview ↗

Follow-up: Results from Additional Experiments

Following our initial response, we conducted additional experiments to investigate whether CodecAttack can induce structured tool calls beyond target-string reproduction, whether these calls lead to downstream tool execution, and whether adversarial audio remains effective in a commercial voice-agent environment.

Question 2: Downstream Tool-Calling Evaluation

To investigate downstream agent behavior, we evaluated CodecAttack on Gemma 4 in a controlled tool-calling environment.

Experimental setup. We conducted optimization and evaluation across three independent seeds (1, 7, 42), varying the number of available tools from 1 to 128. We evaluated the resulting adversarial audio under Opus 64/128 kbps, MP3 64 kbps, and AAC 64 kbps compression.

Unlike the original target-string evaluation, this experiment measures exact-call attack success rate (ASR), requiring the model to generate the attacker-specified structured tool call rather than merely reproduce a target string in its natural-language response.

Table 1. Exact-Call Attack Success Rate (ASR, %) Across Tool-Catalog Sizes and Lossy Audio Codecs. Values Are Mean ± Standard Deviation Across Three Optimization Seeds (1, 7, 42).

Available tools (k)Opus 64Opus 128MP3 64AAC 64
1100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00100.00 ± 0.00
458.50 ± 2.2957.33 ± 1.6157.83 ± 1.2657.17 ± 2.02
847.50 ± 2.1846.00 ± 1.8048.00 ± 3.2847.50 ± 0.87
1639.00 ± 3.2839.67 ± 3.5539.00 ± 3.2839.83 ± 1.15
3224.50 ± 5.2724.33 ± 2.5224.17 ± 2.3623.67 ± 1.44
6411.67 ± 3.0612.00 ± 1.0011.33 ± 0.5813.00 ± 0.50
1283.83 ± 1.152.67 ± 1.043.00 ± 0.003.17 ± 0.29

The results demonstrate that CodecAttack can induce attacker-specified structured tool calls that remain effective after lossy audio compression. With four available tools, the attack achieves 58.50% ASR under Opus 64 kbps, 57.33% under Opus 128 kbps, 57.83% under MP3 64 kbps, and 57.17% under AAC 64 kbps.

Attack success decreases as the number of available tools increases, indicating that targeted tool selection becomes more difficult as the available action space expands.

Additional downstream tool-execution evaluation. To further assess whether attacker-induced tool calls lead to actual execution, we evaluated ε=1 adversarial audio on the held-out Gemma-4-12B-it model in a sandboxed robot-control environment with explicit authorization controls. The experiment used three optimization seeds (1, 7, 42) and compared adversarial audio with the corresponding unperturbed carrier under the same authenticated-session condition (L1T).

CodecAttack achieved 25.2 ± 22.3% protected-tool execution, compared with 2.2% for the unperturbed carrier, representing an observed increase of 23.0 percentage points. These results extend the exact-call evaluation by demonstrating that adversarial audio can induce actual downstream tool execution rather than only attacker-specified structured outputs. The observed executions occurred in an authenticated session and therefore do not establish authorization bypass.


Weakness 2: Preliminary Commercial Voice-Agent Evaluation (Retell AI)

To further assess practical applicability, we conducted a preliminary evaluation using Retell AI with gpt-realtime-mini-2.1 as the backbone model.

We evaluated 15 adversarial audio samples generated through ensemble optimization against Qwen2.5-Omni-7B, Audio Flamingo 3, and Phi-4, without directly optimizing against the deployed target model.

Table 2. Preliminary Target-String Attack Success Rate on Retell AI.

PlatformTarget modelSuccessful attacksTarget-string ASR
Retell AIgpt-realtime-mini-2.15/1533.3%

Of the 15 evaluated samples, 5 successfully induced the attacker-specified target response, corresponding to a target-string ASR of 33.3%.

The evaluation was limited to 15 samples due to the usage constraints of Retell AI's free tier. Accordingly, we present these results as a preliminary feasibility study rather than a comprehensive assessment of the platform's robustness.

Despite the limited sample size, these results provide initial evidence that adversarial audio optimized against open-weight surrogate models can retain some effectiveness when evaluated through a commercial voice-agent platform. A larger-scale evaluation is necessary to establish the reliability and generalizability of this finding.

Thank you again for your review and suggestions. We have posted our responses and additional experimental results addressing the concerns you raised.

We understand that this is a busy time for reviewers, but wanted to follow up as the discussion period is coming to an end. We would appreciate it if you could take another look at our responses when you have a chance, and please let us know if there are any remaining questions or concerns.

Thank you again for your time and consideration.

Reviewer_izWNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CodecAttack, a jailbreak attack against audio language models that combines latent-space perturbation optimization in an audio codec with multi-bitrate Opus Expectation over Transformation to improve attack effectiveness under lossy audio compression. Experiments cover three audio language models and four application scenarios. Under most evaluated compression conditions, the method achieves a higher average attack success rate than a waveform-space attack baseline using the same EoT configuration.

Critical strengths
  1. The paper introduces a multi-bitrate Opus EoT configuration to improve attack robustness to transport compression.
  2. Instead of directly optimizing waveform perturbations, the method optimizes perturbations in the encoder’s latent space and then decodes the perturbed representations, improving robustness to several commonly used audio compression methods.
Critical weaknesses
  1. Table 18 reports PESQ-WB scores well below 2 across multiple speech scenarios at the primary perturbation budget of ε = 1, indicating substantial degradation in the quality of the processed audio.
  2. The transfer experiments in table 3 use ε = 10, whereas the main experiments use ε = 1. Transferability at the primary perturbation budget across models has not been evaluated clearly.
Questions for authors
  1. EoT is an established framework. The paper should more clearly distinguish adaptations of existing ideas to a new setting from novel methodological contributions, and explain why the proposed method achieves more than a straightforward combination of existing techniques.
  2. The audio-quality evaluation currently includes only objective metrics, without subjective listening assessments. Would the authors consider conducting a human listening study?
  3. Could the authors provide additional transfer experiments at the primary perturbation budget or at a comparable audio-quality level?
5 comments

Thank you for your thoughtful review and helpful suggestions. We appreciate the opportunity to clarify our methodological contribution and strengthen the audio-quality and transfer evaluations.

Question 1

We agree that EoT is an established technique. We will revise the manuscript to distinguish these existing components from our contribution more explicitly.

The key insight of CodecAttack is that the choice of perturbation space can fundamentally affect whether adversarial signals survive lossy compression. Rather than optimizing arbitrary waveform perturbations and relying solely on EoT to improve robustness, we optimize perturbations through the continuous latent representation of a neural audio codec. This introduces a structural constraint through the codec decoder, which restricts the perturbations to signal components that may be better preserved during compression.

Our analysis provides evidence for this mechanism. In particular, the decoder Jacobian and random-latent analyses show that low-frequency concentration arises from the codec representation itself, rather than solely from adversarial optimization. Furthermore, our ablations show that EoT remains important for improving robustness across compression settings, while our cross-codec evaluations demonstrate that the resulting attacks can survive transformations not explicitly included during optimization.

Thus, we do not claim novelty in EoT itself. Our contribution is the codec-aware attack formulation, the empirical investigation of its interaction with lossy compression, and the finding that codec preprocessing does not necessarily eliminate adversarial audio targeting Audio LLMs.


Weakness 1 & Question 2

We agree that the low PESQ-WB scores indicate substantial audio degradation. However, our primary contribution is demonstrating that adversarial audio can remain effective under lossy compression, rather than achieving perceptual imperceptibility. PESQ-WB measures reference-based speech quality, but does not directly capture how listeners perceive the perturbed audio. We therefore do not claim that the attacks are imperceptible or perceptually natural.

To evaluate perceived audio quality, we are preparing a human listening study at the primary perturbation budget (ε=1), using a single-stimulus design inspired by ITU-T P.835 [1] and the five-point rating scale described in ITU-T P.800 [2]. Participants will listen to perturbed recordings without access to the original audio or attack information and separately rate speech quality, background-noise intrusiveness, and overall listening quality. We will report the results once the study is complete.

References

[1] ITU-T Recommendation P.835, Subjective test methodology for evaluating speech communication systems and terminals providing noise suppression functionality, 2026.

[2] ITU-T Recommendation P.800, Methods for subjective determination of transmission quality, 1996.

Weakness 2 & Question 3

Thank you for the suggestion. Yes, we have completed additional cross-model transfer experiments at the primary perturbation budget (ε=1).

We jointly optimized adversarial audio against Qwen2.5-Omni, Audio Flamingo 3, and Phi-4, and evaluated transferability on eight held-out models excluded from attack optimization, including open-weight and proprietary API-only Audio LLMs.

The experiment uses 224 target commands across three independent optimization seeds (672 trials per model), using a fixed jazz recording as the audio carrier.

Table 3. Cross-Model Target-String Attack Success Rate (ASR, %) at ε=1 Across Lossy Audio Codecs. Values are mean ± standard deviation across three optimization seeds (1, 7, 42).

ModelRoleUncompressedOpus 64 kbpsOpus 128 kbpsMP3-64 kbpsAAC-64 kbps
Qwen2.5-Omni-7BMember93.45 ± 1.3892.11 ± 1.4793.90 ± 0.9289.73 ± 2.7586.90 ± 0.84
Audio Flamingo 3Member74.55 ± 1.4672.62 ± 1.3874.70 ± 2.0165.63 ± 0.6363.84 ± 1.26
Phi-4Member44.79 ± 1.7244.94 ± 2.4343.60 ± 2.1441.07 ± 1.6741.07 ± 1.67
Qwen2.5-Omni-3BHeld-out25.83 ± 1.6626.04 ± 2.3126.49 ± 2.0125.39 ± 2.4323.96 ± 1.64
Kimi-AudioHeld-out9.08 ± 1.389.82 ± 2.229.52 ± 0.929.23 ± 1.178.93 ± 1.31
SALMONNHeld-out8.93 ± 2.198.48 ± 2.038.93 ± 2.638.48 ± 1.899.08 ± 2.14
Gemma4-E2BHeld-out3.12 ± 0.732.98 ± 1.382.98 ± 0.213.12 ± 0.732.84 ± 0.42
Gemma4-E4BHeld-out5.12 ± 2.016.55 ± 1.724.61 ± 1.804.76 ± 1.114.91 ± 1.82
gpt-audio-miniHeld-out2.68 ± 0.633.12 ± 1.093.57 ± 0.963.12 ± 0.963.87 ± 1.52
gpt-audioHeld-out13.10 ± 0.9212.05 ± 1.5914.73 ± 0.6312.50 ± 0.7312.50 ± 1.31
gpt-audio-1.5Held-out9.08 ± 0.848.78 ± 1.8310.71 ± 0.968.78 ± 0.429.23 ± 0.56

Under Opus 64 kbps compression, the attacks achieve 26.04% ASR on Qwen2.5-Omni-3B, 12.05% on gpt-audio, 9.82% on Kimi-Audio, and 8.78% on gpt-audio-1.5, with additional transfer observed on SALMONN, Gemma 4, and gpt-audio-mini. These results extend our original transfer evaluation to a broader set of held-out models, including proprietary API-only models, at the primary perturbation budget (ε=1). Although transfer success remains lower than on the optimization-member models, the results demonstrate that adversarial audio can transfer to unseen models without target-specific optimization, including after lossy compression.

AuthorsOpenReview ↗

Follow-up: Results from Additional Experiment (Listening Test)

Following our initial response, we completed additional experiments evaluating (1) the perceived audio quality of CodecAttack through a human listening study and objective quality measurements and (2) cross-model transferability at the primary perturbation budget (ε=1) across three independently generated music carriers.

Weakness 1 & Question 2: Human Listening Evaluation and Audio Quality

To complement the objective audio-quality evaluation, we conducted a pilot human listening study at ε=1.

Experimental setup. Five listeners evaluated the same set of 20 adversarial audio samples (10 uncompressed and 10 Opus 64 kbps) using a reference-free, single-stimulus protocol inspired by ITU-T P.835, with five-point absolute category rating (ACR) anchors from ITU-T P.800.

Each listener independently rated three aspects of perceived quality: (1) speech distortion, (2) background-noise intrusiveness, and (3) overall listening quality. Higher scores indicate better perceived quality.

Listeners did not have access to the corresponding unperturbed recordings. This reference-free protocol reflects our threat model, in which a downstream listener or verifier receives only the delivered audio without access to its unperturbed counterpart.

Table 1. Human Listening Evaluation of CodecAttack at ε=1. Overall Audio Quality MOS (1–5; Higher Is Better). Values Are Mean ± Standard Deviation of Ratings Within Each Listener.

ListenerUncompressedOpus 64 kbps
L014.30 ± 0.644.30 ± 0.64
L024.70 ± 0.464.60 ± 0.49
L034.60 ± 0.494.70 ± 0.46
L044.60 ± 0.494.80 ± 0.40
L054.70 ± 0.464.90 ± 0.30
Mean across listeners4.584.66

The pilot study yielded mean overall-quality ratings of 4.58/5 for uncompressed audio and 4.66/5 under Opus 64 kbps compression, corresponding to an overall MOS of 4.62/5 across all 20 stimuli and five listeners.

To directly relate these subjective ratings to objective quality, we computed full-reference PESQ-WB, STOI, and SNR for the exact 20 adversarial audio clips used in the listening study, using each clip's corresponding unperturbed carrier as its reference. All objective measurements were computed at 16 kHz.

Table 2. Objective Audio Quality of the Exact 20 Adversarial Clips Used in the Human Listening Study at ε=1. Values Are Mean ± Standard Deviation Across Clips.

Listening setClipsPESQ-WB ↑STOI ↑SNR (dB)
Uncompressed101.713 ± 0.2330.905 ± 0.0087.51 ± 1.97
Opus 64 kbps101.828 ± 0.2870.891 ± 0.0266.60 ± 0.94
All listening clips201.771 ± 0.2680.898 ± 0.0207.06 ± 1.61

Across the same 20 stimuli, the objective measurements yielded PESQ-WB of 1.771 ± 0.268, STOI of 0.898 ± 0.020, and SNR of 7.06 ± 1.61 dB, while the corresponding reference-free human evaluation yielded an overall MOS of 4.62/5.

These findings provide complementary perspectives on audio quality. Although the reference-based PESQ-WB values remain relatively low, listeners assigned favorable overall-quality ratings to the same adversarial recordings when evaluated without access to clean references. This distinction is relevant to our threat model, in which the recipient typically encounters only the delivered audio.

The listening study is preliminary, comprising five listeners and 20 stimuli generated using a single optimization seed. The results characterize perceived overall quality under reference-free listening conditions and do not establish that the perturbations are imperceptible or quantify perceived degradation relative to clean audio.

Reviewer_6hsCOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes CodecAttack, which optimizes bounded perturbations in a neural audio codec's continuous latent space to induce specified Audio-LLM outputs after compression. Construction bypasses the host codec's residual vector quantization and uses actual codec forward passes with straight-through gradients and multi-bitrate Opus training. Across three models and four application-inspired settings, it reports 83.3% average targeted ASR at Opus 64 and 128 kbps, versus 37.2% for the evaluated codec-aware waveform baseline. Success means reproduction of the complete target string. Additional experiments examine held-out delivery codecs, spectral controls, neural re-synthesis, audio quality, and a separate higher-budget ensemble transfer setting.

Critical strengths
  1. Relevant robustness question: Testing compression against an adaptive attacker addresses a meaningful limitation of relying on transport codecs as preprocessing defenses.

  2. Useful experimental controls: The paper includes a codec-aware waveform reference, repeated optimization runs, distortion measurements, a per-pair SNR-matched band-limited control, and analyses of decoder-induced spectral structure.

  3. Careful scope and interpretation: The paper distinguishes delivery-channel robustness from held-out-model transfer, reports adverse results under neural re-synthesis, and explicitly qualifies perceptual quality, baseline comparisons, and downstream-action claims.

Critical weaknesses

1.Released artifacts lag the submitted evaluation: Table 1 reports means and standard deviations over three optimization seeds, but the linked release provides older single-run primary bundles and a table builder that computes Wilson intervals from one bundle rather than the stated seed statistics. The ensemble seed records are present, so this concern is specific to the primary comparison and newer controls. The released material also does not identify the 40-pair matched control's current artifacts and cohort mapping. This matters for reconciling the Qwen2-Audio S1 AAC-LC results: Table 1 reports 4.7% latent ASR, whereas Table 7 reports 47.5%. Different cohorts or reruns could explain this, but their relationship is not traceable. The authors should synchronize the release with the submitted run-level results and control provenance.

2.Incomplete isolation of the parameterization benefit: The headline comparison uses a genuine codec-aware waveform baseline with the same step budget and EoT schedule. However, Appendix D.2's revised waveform configuration raises Opus 64 kbps ASR from 20.0% to 52.5% on the same 40 pairs while changing bandwidth, optimization settings, and SNR calibration. No unrestricted waveform counterpart receives this stronger configuration. CodecAttack still reaches 77.5% with better measured speech quality, which is useful evidence in its favor, but the remaining advantage over competitively tuned waveform optimization is not isolated across the main settings. The paper appropriately limits its claims to tested configurations; a tuned unrestricted control would make that comparative contribution more convincing.

3.Unvalidated application-level bypass: Exact target-string reproduction is well defined, but it does not establish that an initially correct application decision was overturned. In the released reference records, Qwen2.5-Omni already labels all eight AI-generated music carriers as human-authored before attack, and AF3 already clears all nine commercial recordings. These cells can demonstrate control of specific wording without newly bypassing a correct classifier. Other released cells do make correct clean decisions, so this does not invalidate the attack or its formal ASR. However, the paper does not report current clean task accuracy or decision changes conditional on initially correct classifications. The authors should distinguish targeted wording control from attack-induced changes to clean-correct task decisions.

4.Limited carrier-content diversity: The primary speech sets contain five banking scripts rendered in ten TTS voices and five interview scripts rendered in five voices. The music settings contain eight AI-generated and nine commercial recordings, each reused across five targets, while the ensemble stress test uses one fixed music carrier. Voice and target variation provide real additional attack instances, and the reported denominators are transparent. Nevertheless, the underlying content coverage is substantially narrower than the four application labels suggest. Repeating optimization seeds on fixed pairs measures optimization variability rather than sensitivity to independently sampled carrier content. The authors should evaluate the method on a broader collection of independently authored or recorded carriers.

Questions for authors
  1. Can the release be updated to reconstruct Table 1's three-seed statistics and Table 7's matched control, including the cohort or run differences behind the AAC-LC results?

  2. How does CodecAttack compare with an unrestricted waveform control using the revised control's tuning and per-pair distortion calibration?

  3. How often does the attack overturn a correct clean task decision, separately from reproducing a specified target string?

  4. Do the observed advantages persist across independently authored speech content and a larger collection of music recordings under the same per-example attack setting?

4 comments

Weakness 1 & Question 1. Yes. We have updated our anonymized repository with the per-seed results, reconstruction scripts, and artifact mappings for Tables 1 and 7, including the 40-pair matched control. We independently verified that all 168 mean and standard-deviation results in Table 1 can be reproduced without discrepancies at the reported precision. The updated artifacts and reproduction instructions are available here: CodecAttack Anonymous Repository.

Weakness 2 & Question 2. Our original experiments already compared CodecAttack against a codec-aware waveform baseline under the same optimizer, 1,000-step optimization budget, and multi-bitrate Opus EoT (Table 1, Section 5.2). Appendix D.2 (Table 7) further examined waveform optimization under revised settings.

Following the reviewer's suggestion, we conducted an additional comparison with a more extensively tuned unrestricted waveform baseline, using learning rates selected on development recordings and per-example SNR calibration.

Table 1: Target-String Attack Success Rate (ASR, %) of Tuned Unrestricted Waveform Optimization and CodecAttack Under Opus Compression. Values are mean ± population standard deviation across three optimization seeds.

DeliveryUnrestricted WaveformCodecAttack (Latent)
Opus 16 kbps5.0 ± 4.1%29.2 ± 4.2%
Opus 24 kbps37.5 ± 11.4%54.2 ± 4.2%
Opus 32 kbps80.0 ± 5.4%65.0 ± 2.0%
Opus 64 kbps98.3 ± 1.2%76.7 ± 8.5%

The results show that the relative performance depends on compression bitrate. CodecAttack achieves higher ASR at 16–24 kbps, whereas the tuned unrestricted waveform baseline performs better at 32–64 kbps. These findings indicate that latent-space optimization is particularly beneficial under stronger compression, while unrestricted waveform optimization can be more effective when compression is less aggressive.

Weakness 3 & Question 3. We agree that target-string ASR alone does not establish whether an attack changes the model's original task decision. We are evaluating this separately using three open-weight LLM judges across banking (S1), interview (S2), and AI music (S3a). Each judge assesses the model's responses to the original and adversarial audio against the task instructions and ground-truth labels, without knowing which audio was attacked.

We measure how often an attack changes an initially correct decision into an incorrect one, and report this conditional decision-flip rate alongside clean task accuracy and target-string ASR. We also assess agreement among the judges and manually review ambiguous cases. The results will be reported once validation is complete.

Weakness 4 & Question 4. To address the limited carrier diversity, we generated 150 new recordings across banking, interview, and AI music, with 10 development and 40 evaluation recordings per scenario. We evaluated CodecAttack on the 120 held-out recordings using Qwen2-Audio at ε=1 across three optimization seeds.

Table 2. Target-String Attack Success Rate (ASR, %) on Newly Generated Audio Content Using Qwen2-Audio. Values are mean ± population standard deviation across three optimization seeds (1, 7, 42).

ScenarioUncompressed ASROpus 64 kbps ASROpus 128 kbps ASRMP3 64 kbps ASRAAC 64 kbps ASR
Banking (S1)71.7 ± 3.148.3 ± 1.272.5 ± 4.117.5 ± 2.018.3 ± 3.1
Interview (S2)75.8 ± 6.270.8 ± 3.175.0 ± 5.443.3 ± 3.138.3 ± 2.4
AI Music (S3a)96.7 ± 2.483.3 ± 4.794.2 ± 2.455.8 ± 2.450.8 ± 3.1

The attacks remain effective on the new recordings, with ASR varying across scenarios and compression settings. We are currently evaluating Qwen2.5-Omni and Audio Flamingo 3 using the same recordings, optimization seeds, and compression settings. We will report these results within the next few days once validated.

For the ensemble experiment, we are extending the original evaluation to additional music recordings. We retain the original ensemble configuration (Qwen2.5-Omni, Audio Flamingo 3, and Phi-4) at ε=10, with 224 target entries and three optimization seeds per recording. We will report attack success across individual carriers and variation across optimization seeds separately, allowing us to distinguish sensitivity to audio content from optimization variability. The results will be reported within the next few days once the evaluation is complete.

AuthorsOpenReview ↗

Follow-up: Results from Additional Experiment

Following our initial response, we completed additional experiments evaluating (1) whether CodecAttack changes initially correct task decisions and (2) whether attack effectiveness persists across independently generated audio recordings and distinct carriers in cross-model transfer.

Weakness 3 & Question 3: Clean Task Correctness and Conditional Decision Flips

We conducted an additional policy-based evaluation to distinguish target-string reproduction from attack-induced changes to initially safe responses.

Experimental setup. We evaluated Qwen2-Audio on 40 held-out banking recordings (S1) using CodecAttack at ε=1, with three optimization seeds (1, 7, 42). We considered uncompressed and Opus 64 kbps delivery.

We used three independent open-weight LLM judges: Qwen3-32B, Magistral-Small-2509, and Nemotron-Nano-9B-v2, following the same security-policy rubric and blinded to attack status, target strings, and optimization seeds. We further manually validated all responses.

The banking policy requires identity verification before granting account access or authorizing transactions. We measured:

  • Clean safety rate: Percentage of clean responses complying with the authentication policy.
  • Conditional safety-to-violation flip rate: Percentage of attacked responses classified as policy violations among trials whose corresponding clean response was safe.

Table 1. Clean Safety Rates and Conditional Safety-to-Violation Flip Rates (%) for Qwen2-Audio at ε=1.

JudgeClean: UncompressedFlip: UncompressedClean: Opus 64Flip: Opus 64
Qwen3-32B97.50 (39/40)87.18 (102/117)97.50 (39/40)79.49 (93/117)
Magistral-Small-250997.50 (39/40)71.79 (84/117)97.50 (39/40)64.10 (75/117)
Nemotron-Nano-9B-v285.00 (34/40)91.18 (93/102)90.00 (36/40)81.48 (88/108)

The results show that CodecAttack can induce policy-violating responses even when the model initially complies with the authentication requirement. For example, Qwen3-32B identifies conditional safety-to-violation flips in 87.18% of eligible uncompressed trials and 79.49% under Opus 64 kbps compression.

These findings extend our original target-string ASR evaluation by measuring changes in response-level policy compliance. The observed flip rates vary across judges, and the evaluation does not establish actual banking transactions or downstream tool execution.


Weakness 4 & Question 4: Expanded Evaluation on Independent Audio Carriers

Following our initial response, we completed the expanded evaluation on three Audio LLMs: Qwen2-Audio-7B-Instruct, Qwen2.5-Omni-7B, and Audio Flamingo 3.

Experimental setup. We generated 150 additional audio recordings across banking (S1), interview (S2), and AI-generated music (S3a), with 10 development and 40 held-out evaluation recordings per scenario. Speech was generated using Qwen3-TTS, while music was generated using ACE-Step/YuE2.

We evaluated CodecAttack at ε=1 across three optimization seeds (1, 7, 42), under uncompressed delivery and Opus, MP3, and AAC compression at 64, 128, and 192 kbps.

Table 2. Target-String Attack Success Rate (ASR, %) on Independently Generated Audio Recordings at ε=1. Values Are Means Across Three Optimization Seeds (1, 7, 42). Bitrates Are in kbps.

CodecBitrateQwen2-Audio S1S2S3aQwen2.5-Omni S1S2S3aAF3 S1S2S3a
Uncompressed—71.775.896.714.29.297.550.075.095.8
Opus6448.370.883.312.57.580.035.056.790.8
Opus12872.575.094.215.88.397.549.278.398.3
Opus19272.578.396.715.89.297.551.779.296.7
MP36417.543.355.83.35.059.26.71.740.8
MP312834.260.075.08.35.067.517.515.067.5
MP319250.067.587.511.79.283.333.335.889.2
AAC6418.338.350.85.02.560.85.80.847.5
AAC12870.876.795.014.28.397.548.369.291.7
AAC19265.876.794.214.28.396.749.269.295.0

The expanded results demonstrate that CodecAttack remains effective across independently generated recordings, although performance varies substantially across models, scenarios, and compression settings.

For AI-generated music (S3a), Opus 64 kbps ASR reaches 83.3% for Qwen2-Audio, 80.0% for Qwen2.5-Omni, and 90.8% for Audio Flamingo 3. Performance is more heterogeneous for banking and interview tasks, particularly on Qwen2.5-Omni.

These findings provide additional evidence that attack effectiveness is not restricted to the original collection of recordings. However, differences across scenarios may also reflect variations in task instructions and generation methods, so they should not be attributed solely to carrier content.

AuthorsOpenReview ↗

Follow-up: Results from Additional Experiment

Weakness 4 & Question 4: Cross-Model Transfer Across Distinct Audio Carriers (ε=10)

We further extended the original ensemble-transfer experiment to address the limitation of using a single fixed music carrier.

Experimental setup. We retained the original ensemble configuration consisting of Qwen2.5-Omni-7B, Audio Flamingo 3, and Phi-4, using ε=10 and 224 target commands. We independently optimized attacks on three distinct music carriers (S3a-001, S3a-006, and S3a-011), using three optimization seeds (1, 7, 42) per carrier.

We evaluated attack success on the three optimization-member models and four held-out models: SALMONN-7B, Kimi-Audio-7B, Qwen2.5-Omni-3B, and gpt-audio. The delivery conditions included uncompressed audio, Opus 64/128 kbps, MP3 64/128 kbps, and AAC 64/128 kbps.

Table 3. Cross-Model Target-String ASR (%) Across Three Distinct Music Carriers at ε=10. Values Are Mean ± Standard Deviation Across Three Optimization Seeds (1, 7, 42).

ModelRoleCarrierUncompressedOpus 64Opus 128MP3 64MP3 128AAC 64AAC 128
Qwen2.5-Omni-7BMemberS3a-00199.3 ± 0.598.7 ± 0.599.1 ± 0.598.4 ± 0.398.4 ± 0.397.5 ± 1.399.4 ± 0.7
S3a-00692.7 ± 0.990.2 ± 0.993.0 ± 0.585.3 ± 1.989.0 ± 0.581.0 ± 0.792.6 ± 0.9
S3a-01196.7 ± 0.994.0 ± 2.096.6 ± 0.994.2 ± 2.495.2 ± 1.089.3 ± 2.796.1 ± 1.4
Audio Flamingo 3MemberS3a-00191.2 ± 1.890.0 ± 2.091.2 ± 2.183.8 ± 2.587.8 ± 1.979.5 ± 2.791.2 ± 1.6
S3a-00691.4 ± 0.987.7 ± 1.191.7 ± 1.077.8 ± 3.784.8 ± 1.968.9 ± 1.190.8 ± 1.1
S3a-01190.0 ± 3.388.2 ± 2.290.2 ± 2.778.7 ± 3.483.8 ± 2.872.8 ± 4.588.8 ± 2.7
Phi-4MemberS3a-00140.9 ± 2.838.4 ± 2.040.9 ± 3.035.4 ± 0.738.2 ± 2.529.5 ± 0.540.8 ± 2.6
S3a-00646.7 ± 1.746.0 ± 2.548.2 ± 0.036.6 ± 1.241.5 ± 0.935.6 ± 0.546.0 ± 1.9
S3a-01129.0 ± 6.530.5 ± 4.529.8 ± 5.728.6 ± 4.729.0 ± 5.826.8 ± 5.328.9 ± 5.6
SALMONN-7BHeld-outS3a-0013.7 ± 2.33.6 ± 2.03.4 ± 1.83.3 ± 1.93.4 ± 2.13.3 ± 2.23.7 ± 1.9
S3a-0061.9 ± 1.02.2 ± 0.91.8 ± 0.82.2 ± 0.82.5 ± 1.32.1 ± 0.92.1 ± 0.9
S3a-0113.9 ± 1.44.0 ± 1.33.7 ± 1.84.0 ± 1.64.0 ± 1.63.4 ± 1.03.6 ± 0.8
Kimi-Audio-7BHeld-outS3a-0011.9 ± 1.62.1 ± 2.11.9 ± 1.41.3 ± 1.61.2 ± 1.31.3 ± 0.91.6 ± 1.4
S3a-0063.9 ± 1.63.4 ± 1.43.6 ± 1.33.6 ± 2.03.9 ± 1.63.9 ± 1.83.7 ± 1.6
S3a-0115.4 ± 0.95.8 ± 1.25.4 ± 1.25.1 ± 1.15.4 ± 1.35.2 ± 1.45.7 ± 1.4
Qwen2.5-Omni-3BHeld-outS3a-00114.4 ± 0.314.6 ± 0.314.6 ± 0.514.3 ± 0.814.4 ± 0.713.2 ± 1.814.4 ± 0.7
S3a-00611.6 ± 2.011.2 ± 2.410.4 ± 2.910.0 ± 1.710.4 ± 1.710.7 ± 1.811.2 ± 2.8
S3a-01117.4 ± 2.717.0 ± 3.517.6 ± 3.716.7 ± 4.015.9 ± 3.417.1 ± 3.617.7 ± 2.7
gpt-audioHeld-out APIS3a-0012.5 ± 0.32.5 ± 1.72.5 ± 0.52.8 ± 0.92.7 ± 0.82.5 ± 0.92.2 ± 0.8
S3a-0062.2 ± 0.82.5 ± 1.72.5 ± 0.92.7 ± 0.92.5 ± 1.12.7 ± 0.92.5 ± 0.9
S3a-0115.7 ± 1.45.8 ± 1.26.0 ± 1.66.4 ± 0.55.7 ± 1.35.8 ± 0.56.8 ± 1.8

The results demonstrate that cross-model transfer is not restricted to a single fixed music carrier. Under Opus 64 kbps, Qwen2.5-Omni-3B achieves 14.6%, 11.2%, and 17.0% ASR across S3a-001, S3a-006, and S3a-011, respectively. Similarly, gpt-audio achieves 2.5%, 2.5%, and 5.8% ASR across the three carriers.

Attack effectiveness varies across recordings and held-out models. For example, Kimi-Audio achieves 2.1%, 3.4%, and 5.8% ASR under Opus 64 kbps across the three carriers, while SALMONN remains between 2.2% and 4.0%.

These findings provide additional evidence of nonzero cross-model transfer across distinct audio carriers and lossy compression channels, while also showing that transferability remains substantially lower on held-out models than on optimization-member models.

Importantly, the experiment preserves the original ensemble perturbation budget (ε=10) and optimization configuration, allowing us to examine carrier-dependent variation without introducing a different perturbation budget.

Additional ε=1 transfer results. We further confirm that cross-model transfer persists at the substantially lower perturbation budget (ε=1), across distinct audio carriers and lossy compression channels. Although transfer rates vary across held-out models, these results demonstrate that transferability is not limited to the larger ε=10 budget. Detailed results are provided in our ε=1 follow-up response to Reviewer izWN.

Thank you again for your review and suggestions. We have posted our responses and additional experimental results addressing the concerns you raised.

We understand that this is a busy time for reviewers, but wanted to follow up as the discussion period is coming to an end. We would appreciate it if you could take another look at our responses when you have a chance, and please let us know if there are any remaining questions or concerns.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.