acceptodds
Under review as a conference paper at ICLR 2027

ActionWhisper: Hijacking What Voice Agents Do, Not Just What They Say

Abstract

Voice agents translate spoken requests into structured tool calls, exposing application state to adversarial audio. An attacker-specified action requires the model to generate a valid call with the intended tool and exact arguments. Revisiting response-level audio attacks reveals that an attack can find a waveform that executes the target action, yet return a failed candidate with lower response loss. Teacher-forced loss evaluates the target sequence under correct prefixes, so it can guide waveform updates without reliably ranking freely generated calls. We propose **ActionWhisper**, which uses this loss for search and execution outcomes for candidate retention. Starting from a bounded speech-difference perturbation, it assigns distinct weights to call-format, tool-name, and argument tokens. A two-phase schedule emphasizes format and tool choice before increasing argument emphasis, keeping the complete call in the objective. At regular intervals, it decodes without target prefixes and evaluates the resulting calls through the parser and executor. Candidate selection prioritizes exact action matching, target-tool invocation, and call validity, using response loss to break ties. Across four Audio LLMs and two agent frameworks on executable cases, ActionWhisper achieves the highest target-tool invocation and precise-action rates in all eight settings under matched perturbation budgets and step caps, raising macro-average precise-action rate to , versus for the strongest baseline, Target-Text CE. Code is available at https://anonymous.4open.science/r/actionwhisper-review/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agent Security, Audio LLM, Voice Agent

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RKvVOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the security risks of tool calls in voice agents. The authors identify a limitation of relying on teacher-forced loss for optimization and candidate selection. A lower loss does not always lead to successful execution and may cause successful attack candidates to be discarded. To address this issue, the authors propose ActionWhisper, which combines speech-difference initialization, two-phase argument curriculum, and executor-grounded candidate selection to improve attack success rates. The authors also introduce Action-Suite-120, a benchmark containing 120 executable attack cases. Experiments on four Audio LLMs and two agent frameworks compare ActionWhisper with existing baselines.

Critical strengths
  1. This paper highlights the risk of audio adversarial attacks causing incorrect tool calls in voice agents. Such risks may be overlooked when evaluating only the model's generated responses.This is a relevant security concern worth investigating. The proposed execution-grounded candidate selection strategy also provides a practical way to evaluate and retain successful attack candidates.

  2. The paper identifies a limitation of relying solely on teacher-forced loss for adversarial audio optimization. A lower loss does not necessarily correspond to successful execution of the target action and may even cause successful attack candidates to be discarded during selection. This observation is interesting and provides useful insight into the mismatch between optimization objectives and actual execution outcomes.

Critical weaknesses
  1. Lack of innovation. The paper focuses on audio adversarial attacks that induce incorrect tool calls in voice agents. Although this is an important problem, the paper does not provide enough evidence to show how common or serious this issue is in existing attacks. The proposed method mainly combines existing techniques, including gradient-based optimization, staged loss weighting, and executor-grounded candidate selection. It appears to be an engineering improvement for a specific attack scenario rather than a fundamentally new method.

  2. Limited experimental scale. The evaluation includes only 120 attack cases, with limited diversity in tasks and argument combinations. This makes it difficult to assess how well the findings generalize to other scenarios. In addition, the main experiments use only a single random seed. The paper does not report repeated trials, confidence intervals, or statistical significance tests.

Questions for authors
  1. How common is the loss-execution mismatch in existing attacks? The paper argues that where a lower teacher-forced loss leads to unsuccessful tool execution. However, it is unclear how frequently this occurs across different models and attack methods. Could the authors provide quantitative evidence of this phenomenon and its impact on attack success rates?

  2. How much of the improvement comes from executor-grounded selection? The main baselines do not use the same candidate selection strategy as ActionWhisper. Would ActionWhisper still outperform these methods if all baselines used the same strategy as ActionWhisper?

  3. The evaluation uses 120 attack cases and a single random seed. Could the authors provide results across multiple random seeds or statistical significance analysis?

1 comment

Dear Reviewer RKvV,

We are writing to follow up on our response to your three questions. In brief:

  • Q1: Loss-based selection discards 15.8%–25.0% of the exact successes that existing attacks had already found, and the mismatch appears in all eight settings.
  • Q2: With the same executor-grounded selection, ActionWhisper leads Target-Text CE in seven of eight settings.
  • Q3: Across three seeds, ActionWhisper stays ahead on every seed (38.3 ± 0.8 vs. 36.4 ± 2.1 PA).

We would value your view on whether this evidence is sufficient, and we are happy to provide anything further you would find helpful.

Best regards, The Authors

Reviewer_Ak7zOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies audio-to-action hijacking, where bounded perturbations to benign speech induce a voice agent to execute an attacker-specified tool call with exact arguments. It identifies cases where teacher-forced target loss favors failed candidates over successful ones and proposes ActionWhisper, which combines speech-difference initialization, staged weighting of call-format, tool-name, and argument tokens, and execution-based candidate selection. The authors introduce Action-Suite-120, comprising 120 sandboxed cases across six risk categories, and evaluate four audio LLMs under two agent frameworks. ActionWhisper achieves a macro-average precise-action rate of 28.3%, compared with 22.0% for Target-Text CE. Additional experiments examine component contributions, carrier variation, additive noise, audio purification, and computational cost.

Critical strengths
  • The paper addresses a meaningful security problem by examining whether adversarial audio can induce specific tool calls with exact arguments. Checking actual execution makes the evaluation more informative.

  • The observation that loss-based selection can discard an already successful attack is useful and provides a clear motivation for execution-based candidate selection.

Critical weaknesses
  • The novelty appears limited. Prior work has studied adversarial audio and demonstrated unauthorized tool execution. Methodologically, ActionWhisper combines target-dependent initialization and execution-based selection within an existing gradient-based attack framework.

  • The main comparison changes both optimization and candidate selection. Although the paper includes component ablations, it remains unclear how much advantage the proposed search strategy offers over existing attacks equipped with the same execution-based selection.

  • Failure to execute the exact target action does not necessarily mean that the agent resisted the attack. It may produce invalid calls, incorrect arguments, or unrelated responses while also failing the user's original task. The paper does not sufficiently distinguish preserved legitimate behavior from disruption that falls short of the attacker's exact objective.

  • Attack success also depends on the model's underlying tool-use ability. The low clean-audio call-validity rates in some settings raise the possibility that low attack success partly reflects difficulty producing executable calls, rather than resistance to the perturbation. This complicates the interpretation of differences across models.

Questions for authors

Among attacks that fail to achieve the exact target action, how often does the agent still complete the user's original task? The existing failure breakdown is useful, but it primarily measures progress toward the attacker's goal. Low PA indicates difficulty achieving precise control, but does not by itself establish that the agent resisted disruption.

1 comment

Dear Reviewer Ak7z,

Thank you again for your review. Your question about what happens when the exact action fails led to an analysis that is now part of the paper: among PA failures, the agent completes the user's original task in only 17.6% of cases, and 22.5% of the full suite preserves clean behavior with no unauthorized action. The explicit-target-speech diagnostic (W4) and the shared-selection comparison (W2) are in the revision as well.

Could you let us know whether these analyses resolve your concerns? We would be glad to clarify anything that remains unclear while the discussion is open.

Best regards, The Authors

Reviewer_N13uOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ActionWhisper, a white-box adversarial audio attack designed to hijack the tool execution of voice agents. Unlike conventional audio attacks that primarily manipulate model-generated textual responses, this work aims to induce an agent to execute an attacker-specified tool call with exact arguments. The proposed method combines speech-difference initialization, a two-phase optimization strategy that assigns different weights to call-format, tool-name, and argument tokens, and execution-grounded checkpoint selection. The authors introduce Action-Suite-120 and evaluate the method across four Audio LLMs and two agent frameworks. Experimental results show that ActionWhisper improves the macro-average precise-action rate from 22.0% for the strongest baseline to 28.3%, with additional evaluations covering component ablations, additive noise, and attack-aware purification.

Critical strengths

Interesting and practically relevant problem. The paper investigates adversarial audio attacks against executable agent actions rather than only generated responses.

Execution-aware optimization with promising results. The proposed method recognizes the mismatch between minimizing response-level loss and achieving successful tool execution. The experimental results across multiple models and agent frameworks demonstrate improvements over existing approaches.

Critical weaknesses

Limited practical feasibility under the assumed threat model. The attack requires white-box model access and digitally supplied adversarial waveforms. It remains unclear how the adversarial audio would be delivered in realistic voice-agent applications. The robustness evaluation also lacks physical acoustic experiments.

Insufficient differentiation from existing methods. While action-level optimization is interesting, the underlying approach largely builds on existing adversarial waveform optimization, with additional token weighting and execution-based checkpoint selection. The paper does not fully establish the fundamental methodological novelty, and the baseline comparisons could better isolate the benefits of these components.

Questions for authors

How would an attacker practically launch this attack? Section 4.1 assumes that the attacker can modify the benign audio waveform before it is supplied to the voice agent. However, it is unclear how this assumption translates to a realistic attack scenario. Does the attacker need to replace the entire benign audio with the adversarial waveform, or can the attacker play the optimized perturbation through an external speaker while the legitimate user is speaking? If the latter, how would the attacker handle synchronization and differences in speech timing? I encourage the authors to clarify the attacker's capabilities and provide a concrete deployment scenario.

How robust is the attack under realistic acoustic conditions? I appreciate the additive-noise experiments in Section 5.5 and Appendix I. However, the precise-action rate decreases substantially as the channel noise increases. More importantly, the experiments consider Gaussian noise rather than physical acoustic propagation. Could the authors evaluate the attack through speaker-to-microphone playback, considering reverberation, background speech, device differences, and microphone distance? Such an evaluation would substantially strengthen the practical implications of this work.

Why is the evaluation restricted to relatively small, open-source Audio LLMs under white-box access? The paper motivates the problem using practical voice-agent systems, including commercial deployments. However, the proposed optimization requires model gradients, which are generally unavailable for proprietary voice-agent services. Could the authors clarify whether adversarial waveforms generated using one model can transfer to another model without additional optimization? It would also be valuable to understand whether the method remains effective against larger and more capable Audio LLMs. This is important for assessing the generalizability of the proposed threat.

What is the fundamental technical distinction between ActionWhisper and existing adversarial audio attacks? Prior work, such as AudioHijack, has already demonstrated auditory prompt injection and unauthorized tool execution. The proposed method introduces structured call optimization and execution-grounded candidate selection, but I am not fully convinced that these constitute a fundamentally new attack methodology. In particular, retaining candidates that successfully achieve the attack objective appears to be a relatively straightforward modification. Could the authors more clearly explain the technical insights that differentiate ActionWhisper from existing approaches? Additionally, could the authors provide a full-suite comparison where the strongest baselines use the same execution-grounded checkpoint-selection strategy? This would help isolate the contribution of the proposed optimization method from the benefits of improved candidate selection.

1 comment

Dear Reviewer N13u,

Thank you for the detailed questions on deployment, which shaped much of our revision. Since your review draws on direct experience with this topic, we would particularly value your assessment of the new evidence: the digital-ingestion delivery path (Q1), the simulated acoustic conditions (Q2), zero-shot cross-model transfer (Q3), and the full-suite comparison with selection shared across methods (Q4).

If any of these still falls short of what you had in mind, please let us know and we will follow up promptly.

Best regards, The Authors

Comments on the paper

2 comments

Dear Reviewers and Area Chair,

We thank all three reviewers for their careful reading. We are encouraged that every review recognizes action-level hijacking of voice agents as a relevant security problem and the loss–execution mismatch as a useful observation. Our individual responses and the revised PDF were posted on Nov 11. For convenience, this note groups the concerns by theme and gives the key evidence for each.

Concerns shared across reviews

1. Comparison under the same executor-grounded selection (N13u Q4, Ak7z W2, RKvV Q2). We equipped Target-Text CE with our selector in all eight model–framework settings, and all four baselines with it on Qwen2-Audio/LangChain. Target-Text CE rises from 22.0 to 26.8 macro-average PA. ActionWhisper reaches 28.3, leads in seven of eight settings, and stays ahead on each of three seeds on Qwen2-Audio/LangChain.

2. Technical distinction from prior attacks (N13u W2, Ak7z W1, RKvV W1). The method follows from a diagnosis. Teacher-forced loss can reverse the ranking of freely decoded calls (Lemma 1, Theorem 1), and loss-based selection discards 15.8%–25.0% of the exact successes that existing attacks had already found along their own trajectories. Each component answers one diagnosed bottleneck. With selection held fixed (Table 3), the two-phase schedule contributes 11.7 PA points and speech-difference initialization 7.5.

Reviewer-specific points

  • N13u (delivery, acoustics, transfer). Digital audio ingestion gives a delivery path that needs no synchronization with a live speaker. Replaying the perturbation on an unseen carrier keeps 37.5% PA, and simulated device response, mild reverberation, and competing speech keep 21.7%–26.7% PA without re-optimization. Zero-shot cross-model transfer is reported for all three other models.
  • Ak7z (task preservation, native tool use). Among PA failures, the agent completes the user's original task in 17.6% of cases, and 22.5% of the full suite preserves clean behavior with no unauthorized action. Under explicit clean target speech, call validity is 75.0%–94.2%, and ActionWhisper reaches 42.7%–60.0% of the explicit-command ceiling on three of the four models.
  • RKvV (prevalence, statistics). The mismatch appears in every one of the eight settings, where loss-based selection discards 17.9% of the exact successes Target-Text CE had found. Three-seed results are in our response and the revision.

We remain available for the rest of the discussion period. If any point is still open, we would be glad to hear it early enough to run further experiments or revise the manuscript again.

Best regards, The Authors

AuthorsOpenReview ↗

Summary of Rebuttal and Revision

Dear AC and Reviewers,

Thank you for your careful reviews and constructive suggestions. As the author–reviewer discussion closes, we have not yet received follow-up comments, so we summarize below how the revision addresses each concern, for convenience in the upcoming discussion. All points are supported by new experiments and analyses in the revised manuscript.

  • ActionWhisper stays ahead when every method uses the same executor-grounded selection. With our selector, Target-Text CE rises from 22.0 to 26.8 macro-average PA; ActionWhisper reaches 28.3 and leads in seven of eight settings. On Qwen2-Audio/LangChain it leads all four baselines equipped with the selector, and stays ahead on each of three seeds (38.3 ± 0.8 vs. 36.4 ± 2.1). (N13u Q4, Ak7z W2, RKvV Q2–Q3)
  • The loss–execution mismatch is common in existing attacks. Loss-based selection discards 15.8%–25.0% of the exact successes that Target-Text CE, AdvWave, and AudioHijack had already found, and the mismatch appears in all eight settings. (RKvV Q1)
  • Each component answers a diagnosed bottleneck. Teacher-forced loss can reverse the ranking of freely decoded calls (Lemma 1, Theorem 1). With selection held fixed, the two-phase schedule contributes 11.7 PA points and speech-difference initialization 7.5 (Table 3). (N13u W2, Ak7z W1, RKvV W1)
  • The attack needs no synchronization with a live speaker and persists without re-optimization. Digital audio ingestion gives a direct delivery path. The perturbation keeps 37.5% PA on an unseen carrier and 21.7%–26.7% PA under simulated device response, mild reverberation, and competing speech. Zero-shot transfer yields exact-action successes on all three other models, where clean and noise inputs yield none. (N13u Q1–Q3)
  • Missing the exact action rarely means the agent was unaffected. Among PA failures, the agent completes the user's original task in only 17.6% of cases, and 22.5% of the full suite preserves clean behavior with no unauthorized action. (Ak7z W3, Q1)
  • Differences across models are not explained by tool-use ability alone. Under explicit clean target speech, call validity is 75.0%–94.2%, and ActionWhisper reaches 42.7%–60.0% of the explicit-command ceiling on three of the four models. (Ak7z W4)

Full tables are in our individual responses and the revised PDF. We remain available until the discussion closes and would be glad to answer any remaining questions. Thank you again for your time.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.