acceptodds
Under review as a conference paper at ICLR 2027

Crescendo-Live: Multi-Turn Jailbreaks of Full-Duplex Speech Models in Live Calls and Their Defense with Daimonion

Abstract

Full-duplex spoken dialogue models listen and speak at the same time, aiming at seamless, natural voice conversation. The latency constraint limits the model's ability to reason about what it hears, which leaves it exposed to multi-turn jailbreaks such as Crescendo: the attacker opens with a benign question and escalates gradually, building each request on the model's previous response, so the harmful goal is spread across turns and becomes more implicit. Yet the safety of speech models is still evaluated on single harmful requests or pre-scripted dialogues. We present crescendo-Live, which evaluates full-duplex models in live multi-turn conversation: an LLM-driven user simulator adapts every turn to the model's response and can interrupt the model while it is speaking. Multi-turn escalation jailbreaks every model we evaluate, even those that are robust to direct requests, and attack success grows with the number of turns. Our ablations also find that talking over the model leads to an even higher attack success rate, raising a safety concern unique to full-duplex models. Aligning the speech model itself against such attacks is hard: one hidden goal can be delivered in many different ways, and full-duplex training data is expensive to synthesize. We instead propose Daimonion, which pairs a full-duplex model with an asynchronous LLM safety monitor. The monitor follows the whole conversation, and when it detects a harmful intent, it injects a text instruction into the frontend full-duplex model, which is fine-tuned to turn such instructions into a spoken refusal while keeping its natural conversational ability. Daimonion reduces attack success to below 3% at every conversation length while retaining the model's original turn-taking ability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
full-duplex spoken dialogue models, speech language models, multi-turn jailbreak, Crescendo, red teaming, user simulation, AI safety, safety monitor

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_DrxGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the vulnerability of full-duplex speech models to adaptive, multi-turn jailbreaks in live calls. It introduces crescendo-Live, a benchmark that extends Crescendo to spoken interaction: an LLM attacker escalates its requests based on the model's actual spoken replies and can interrupt the model mid-speech. On three full-duplex models and one half-duplex reference, attack success rises with the number of turns, reaching about 67–69% at 8 turns for the full-duplex models, and is higher when the attacker interrupts. The authors then propose DAIMONION, which pairs a refusal-fine-tuned speech frontend with an asynchronous LLM safety monitor that injects refusal instructions when it detects harmful intent. On PersonaPlex-7B, it reduces attack success to below 3% across the evaluated turn budgets. Further analyses cover speaking style, monitor latency, and false alarms.

Critical strengths
  1. The paper extends multi-turn jailbreak evaluation to live full-duplex speech, where prior speech-safety evaluations rely on single-turn requests, scripted dialogues, or fixed pre-recorded interruptions. In crescendo-Live the attacker adapts to the model's actual spoken replies and can interrupt it, and the results show vulnerabilities that single-turn testing misses. The interruption ablation on PersonaPlex-7B is an informative finding about a real-time safety risk, which is why I consider the benchmark a valuable contribution.

  2. DAIMONION is a practical defense that keeps safety reasoning off the latency-critical audio path by combining an asynchronous monitor with instruction-conditioned refusal training. It reduces eight-turn attack success on PersonaPlex-7B from 68.9% to 2.9%, and the analysis of detection, false alarms, and alert timing gives useful evidence about deployment trade-offs.

Critical weaknesses
  1. The methodological novelty is limited. Crescendo-Live reuses the published Crescendo attacker prompt, adapted to speech, on top of the τ-Voice simulator (Section 3.2, Appendix A.2), so the novelty lies in extending evaluation to live full-duplex calls with adaptive turn-taking and interruption rather than in a new attack. This is a meaningful extension, but the experiments do not show its incremental benefit. The single-turn setting (k=1) uses a direct request while the multi-turn settings use a different escalation strategy, and a call counts as successful if any model turn is flagged (Section 3.3), so more turns give more chances to succeed even without higher per-turn vulnerability. The rise in Figure 3 therefore cannot be attributed solely to adaptive escalation. More importantly, there is no comparison with a non-adaptive or scripted multi-turn attacker under the same turn budgets and spoken conditions, although adaptivity is the main difference from prior speech-safety evaluations. Without this baseline, the added value of the benchmark is hard to quantify.

  2. The architectural novelty of DAIMONION is limited and the evidence for its independent contribution and cost is insufficient. Pairing a full-duplex frontend with an asynchronous LLM backend has been explored before, e.g., MoshiRAG, so the novelty lies mainly in applying it to safety monitoring and training the frontend to follow injected refusal instructions. The experiments do not isolate these safety-specific components. The main comparison pits the original PersonaPlex-7B against a system that adds refusal fine-tuning and the monitor together, and the fine-tuned frontend without the monitor is not evaluated on the jailbreak benchmark, which is a critical missing baseline. The backend comparisons examine alert decisions on 50 harmful requests and their benign twins, not end-to-end jailbreak results. The monitor also flags 12 of 50 benign twins, a possibly important safety–utility cost, and benign-task performance and latency to a substantive answer are not evaluated for the complete system. It is therefore hard to judge the independent value of the proposed safety mechanisms and the cost of the reported gains.

Questions for authors

Q1. Contribution of the asynchronous safety monitor.

Could the authors report the attack success rates of the refusal-fine-tuned PersonaPlex-7B without the safety monitor, using the same 341 harmful intents and turn budgets k = 1, 4, 8? This comparison would help distinguish improvements from refusal fine-tuning from those specifically attributable to asynchronous monitoring.

Q2. Contribution of adaptive escalation.

Could the authors provide a controlled comparison between crescendo-Live and a non-adaptive or scripted multi-turn attacker under the same turn budgets and interaction conditions? Additionally, reporting the distribution of the first successful attack turn would help distinguish the effects of adaptive escalation from the increased number of opportunities for success.

Q3. Safety–utility and latency trade-offs of the complete defense system.

Could the authors provide additional end-to-end evaluation of DAIMONION on benign multi-turn conversations, including false refusal rates and latency to the first substantive response rather than only the first generated audio? It would also be helpful to quantify how frequently safety alerts arrive after the neutral opener has ended and whether harmful content is emitted before intervention. These results would help establish whether the strong safety gains can be maintained without substantially degrading the usefulness and responsiveness of the speech assistant.

4 comments

We thank the reviewer for the careful review and for recognizing the value of our benchmark. We first answer the three questions, which cover the new experiments, and then respond to the two weaknesses. New and revised text in the updated PDF is marked in blue.

Q1. The fine-tuned speech model without the monitor

We ran this experiment: the fine-tuned speech model with the monitor turned off, on the same 341 intents, with the same attacker and the same grading (Figure 3, Section 5.2). Attack success rate (ASR^HB, %):

kOriginal PersonaPlexFine-tuned model only (monitor off)Daimonion (full system)
135.86.52.1
449.625.51.5
868.932.32.9

Fine-tuning alone improves safety to some extent, mainly for direct requests, because the model saw direct harmful requests with refusals during training. However, the improvement is limited when the harmful intent is spread across several turns: as with all original models, the attack success of the fine-tuned model alone increases with the number of turns. With the monitor, attack success stays low at every turn budget.

Q2. Adaptive vs. scripted attacker, and the first successful turn

For each of the 50 AdvBench intents in our seed set, we use the same attacker model and the same attacker prompt to write a four-turn Crescendo script in advance, without seeing any model reply. The scripted attacker then speaks these four turns in order, with the same voice, and we grade the calls with the same HarmBench classifier. The adaptive attacker is crescendo-Live with the same number of turns (k = 4) on the same intents. Attack success rate (ASR^HB, %; Table 2):

ModelScriptedAdaptive
PersonaPlex-7B38.050.0
MiniCPM-o 4.550.048.0
VoiceChat-11B42.048.0

The adaptive attacker achieves a 12-point higher attack success rate on PersonaPlex, a 6-point higher rate on VoiceChat, and a similar rate on MiniCPM-o. The adaptive attacker can respond to what the model actually says: when the model avoids the question, refuses, or answers only part of it, the attacker can rephrase, take a step back, or follow up. This flexibility increase the ASR.

First successful turn (k = 4, 341 calls per model). Number of calls whose first harmful turn is turn 1, 2, 3, 4:

ModelTurn 1Turn 2Turn 3Turn 4
PersonaPlex-7B7426852
MiniCPM-o 4.53367164
VoiceChat-11B6366257

If more turns only gave more chances, with the same chance of success at every turn, first successes would be most common at turn 1 and less common at each later turn.

Q3. Safety–utility and latency trade-offs of the full system

Utility and latency (Table 1). We compare the full system (with the monitor on) with the original PersonaPlex on LLaMA Questions (300 spoken questions). From the end of each question, we measure when the model starts producing audio and when it says the first word of the actual answer.

MetricPersonaPlexDaimonion
Accuracy (%)57.059.0
First audio (s, median)0.360.45
First answer word (s, median)1.392.32

Daimonion keeps the original accuracy and starts speaking almost as quickly. The actual answer starts 0.9 s later, which is the cost of the short neutral phrase at the start of the reply. The monitor sent no alert on any of the 300 questions, nor on any of the 727 examples of Full-Duplex-Bench v1.0, where the turn-taking results stay close to those of the original model (Appendix A.13).

False refusals in multi-turn conversations. We use the 50 SARSteer benign twins, run with the same Crescendo attacker, which keeps asking for more specific details. With the monitor, the model refused 13 of the 50. We look into the gap and find out most of these are protective requests, such as preventing identity theft or securing a government database, where the simulator start with background or motivation like "what are common way to break..." and follow up with "how can we prevent them". But the monitor will flag the first query as harmful and trigger a refusal.

Alert timing and harmful content before the alert. The monitor decides while the caller is still speaking, so most alerts are ready when the caller stops; for the 23% that are not, the median remaining wait is 0.84 s, shorter than the 2.85 s median neutral phrase (Section 4.2). We also checked the outcome: across the k = 4 and 8 runs, 2,496 model turns received an alert, 824 of them after the model had started its reply, and none was judged harmful (Section 5.2). Harmful turns occur only before the monitor's first alert in a call, that is, when the monitor has not yet detected the attack.

W1. Novelty of crescendo-Live, and whether more turns alone explain the higher attack success

We agree that most components already exist: the attack prompt comes from Crescendo and the call simulation from τ-Voice. Our focus is a threat that is realistic but underexplored: an attacker that adapts each request to the model's spoken replies over multiple turns and can interrupt the model in real time, as a real user can. The closest prior work uses fixed audio: randomly sampled multi-turn (MTR-DuplexBench, Zhang et al., 2025), or one pre-recorded interruption after a single harmful request (DuplexJail, Roh et al., 2026). Our results show that the live setting is more dangerous: Models that resist direct requests (5.6% and 1.8% attack success for MiniCPM-o and VoiceChat at k = 1) are jailbroken in 69.2% and 66.9% of calls at k = 8 (Figure 3); Interrupting raises attack success on PersonaPlex from 29.6% to 49.6% at k = 4 (Table 3); and adapting to the model's replies also raises it (Table 2, Q2).

Regarding whether this rise comes mainly from having more turns, we respectfully disagree. Crescendo splits the harmful request across several harmless-looking turns, so that no single turn reveals the full intent; this, not the number of turns, makes the model easier to jailbreak. Three observations support this:

  • More questions alone do not reduce safety. MTR-DuplexBench asks a separate harmful question in each of 10 rounds, and the refusal rate stays almost the same across rounds (e.g., Moshi: 90.0% in round 1 and 91.0% over rounds 1–10).
  • Successes come late in the conversation. First successes are rare at turn 1 and most common at turns 3–4 (Q2), which is not what more chances alone would produce.
  • The increase is larger than repeated attempts would explain. If each of the k turns were an independent direct request, at most 1 − (1 − p₁)^k of the calls would succeed, where p₁ is the k = 1 rate. For the two models that usually refuse a direct request, this gives 20.6% (MiniCPM-o) and 7.0% (VoiceChat) at k = 4, while the observed rates are 51.0% and 47.2%.

Overall, we appreciate reviewer's effort and hope these additions address the reviewer's concerns. We are happy to provide more details.

Reviewer_DrxGOpenReview ↗

Response

Thanks for the response and the extended results for my question. I will keep my score

Thank you :)

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

The paper studies adaptive multi-turn jailbreaks against full duplex speech models. Crescendo Live combines a Crescendo-style LLM caller with a simulated telephone channel, allowing the caller to adapt its requests to the model's spoken responses and to interrupt ongoing speech. The evaluation covers 341 harmful intents, three full duplex models, and one half duplex reference at budgets of one, four, and eight caller turns. Attack success is measured by transcribing the model's recorded speech and applying the HarmBench classifier to each model turn. A call succeeds if any turn is classified as harmful.

The paper also introduces Daimonion, which combines a refusal trained PersonaPlex frontend with an asynchronous text-LLM monitor that injects internal refusal instructions. The full system substantially reduces observed attack success across all three budgets. Additional experiments examine caller turn-taking and delivery, compare three monitoring backends, measure false alerts on matched benign requests, and assess the trained frontend's turn-taking capabilities.

Critical strengths

1. A consequential evaluation of adaptive spoken attacks

The benchmark addresses an important deployment setting: an adversary can respond to what a speech agent actually says and alter the conversation while the agent is speaking. The distinction from single-request evaluation is empirically consequential. MiniCPM-o and VoiceChat have direct-request attack success rates of 5.6% and 1.8%, respectively, whereas the three full duplex models reach approximately 67-69% at eight caller turns (Sec. 5.2, Figure 3). On PersonaPlex at four turns, enabling the Assertive caller raises attack success from 29.6% for the Patient caller to 49.6% (Table 1). This is useful evidence about the risk introduced by the evaluated interaction policy, even though it does not isolate a purely acoustic mechanism or establish a model safety ranking.

2. A coherent defense with substantial observed gains

Daimonion reduces PersonaPlex's attack success from 35.8%, 49.6%, and 68.9% to 2.1%, 1.5%, and 2.9% at the three budgets (Section 5.2). The conditional refusal training is thoughtful: refusal/control pairs and shared neutral openers aim to make refusal depend on the internal instruction, while interruption augmentation addresses the full-duplex setting. The evaluation scores recorded speech and counts harmful output anywhere in a call, so a later refusal does not automatically erase an earlier harmful turn (Sec. 3.3, App. A.7-A.8). The authors also explicitly examine benign false alerts. These choices make the complete system result promising and provide a useful foundation for further validation.

Critical weaknesses

1. The incremental benefit of the asynchronous monitor is not established

Fig. 3 compares released PersonaPlex with a system that changes both the frontend's training and the monitoring mechanism. The paper does not report harmful call attack success for the identical refusal trained frontend with its monitor disabled. This control is particularly important because the training includes harmful requests and refusals, and Sec. 5.2 explicitly attributes some protection against direct requests to that training: only seven of 65 direct calls without an alert contain harmful content. The subsequent attribution of multi-turn protection to the monitor therefore remains insufficiently tested.

The alternative backend comparison does not resolve this issue. Figure 4 measures whether a backend sends at least one alert. The proposed monitor and THRD both flag 48 of 50 harmful calls, but detection at some point in a call is not equivalent to preventing harmful speech. The paper discusses differences in alert latency and refusal frequency without reporting the corresponding harmful output attack success rates for these backends. Consequently, it is difficult to determine how much practical protection is specifically gained by the proposed monitoring schedule.

To strengthen this contribution, evaluate the already trained frontend with and without monitoring on the existing harmful intents, particularly at four and eight turns. For the backend comparisons already performed, report the same audio based, any turn attack success metric used in the main experiment, alongside alert and refusal rates. These are focused controls that require no new defense architecture. They would establish what monitoring adds beyond training and whether its earlier alerts improve the outcome that matters. The current results support the effectiveness of the complete package, but do not yet adequately support the attribution and comparative advantage of its central defense component.

2. Preservation of useful real-time interaction is insufficiently demonstrated

The claims that safety checks add no response latency and preserve conversational ability are stronger than the available evidence. Section 4.2 deliberately introduces a roughly 3-second neutral opener to allow an alert to arrive before the substantive answer. Figure 2 labels its opener distribution as coming from training, and the reported median is 2.85 seconds. This can plausibly preserve rapid speech onset, but it does not establish unchanged time to a useful answer or refusal. I am not inferring an exact 2.85 second increase over the baseline. That difference requires measuring the baseline's corresponding content onset latency.

The timing comparison also needs to use actual generated responses. Section 5.3 infers that many THRD alerts arrive after answering has begun by comparing their delays with the median opener duration. A training data median does not establish when substantive or harmful content began in each evaluated call. Similarly, the proposed monitor's favorable median remaining wait is encouraging but does not characterize late alert cases. Section 3.2 states that model computation consumes no simulated time, so the relationship between measured monitor processing delays and the simulated speech timeline also needs to be explicit.

There is a related gap in evaluating useful benign interaction with the complete system. The proposed monitor flags 12 of 50 benign twins, or 24%, with false alarms occurring on protective requests such as securing a network or preventing fraud (Figure 4, Section 5.3). The authors appropriately report this cost, but they do not quantify successful benign assistance with the monitor active. Appendix A.12 evaluates turn-taking without a backend, which is a useful frontend check but cannot establish the deployed system's conversational utility.

A bounded way to resolve this concern is to use the existing benign calls and attack recordings to report first audio latency, first useful content latency, benign answer usefulness or completion, and the timing of alerts relative to actual generated content. Clarify how backend and transcription delays enter the simulator and characterize the late-alert cases. The main attack success metric already evaluates recorded speech. The missing evidence concerns the real-time mechanism and the safety-utility tradeoff. These measurements would support a more precise account of the benefit and cost of asynchronous monitoring.

Questions for authors
  • What are the harmful output attack success and refusal rates for the identical fine tuned PersonaPlex frontend with the monitor disabled, especially at four and eight caller turns? How large is the incremental improvement from enabling the monitor?

  • For the proposed monitor, THRD, and the guard gated monitor, what are the audio-based attack success rates on the existing harmful call comparison? How often does harmful speech occur before the first alert, and how often does it occur after an alert has been delivered?

  • Is the 2.85-second opener statistic measured only from training data, or also from generated evaluation responses? How are monitor processing, streaming transcription, and instruction delivery delays mapped onto the simulator clock? Can you report first audio and first useful content latency against the baseline, together with paired alert or content-onset measurements?

  • With the full backend enabled, how often do the existing benign twin calls receive a useful answer or complete their benign objective, versus being unnecessarily refused? Were the realized adaptive conversations checked to remain benign throughout, rather than relying only on the seed intent's label?

  • The paper mentions manual checks of grading against the audio. What were their sample sizes, selection procedure, and agreement or error rates, including for the primary HarmBench attack success decisions? Quantifying this would help interpret the residual defended attack success, which corresponds to only five to ten flagged calls per budget.

8 comments

We thank the reviewer for the detailed review and the concrete suggestions. We answer the five questions below; they also address the two weaknesses, and we comment on the weaknesses at the end. New and revised text in the updated PDF is marked in blue.

Q1. The fine-tuned frontend with the monitor disabled

We ran this control: the same fine-tuned speech model with the monitor turned off, on the same 341 intents, with the same attacker and grading (Figure 3, Section 5.2). Attack success rate (ASR^HB, %):

kOriginal PersonaPlexFine-tuned model only (monitor off)Daimonion (full system)
135.86.52.1
449.625.51.5
868.932.32.9

Enabling the monitor lowers attack success by 4.4, 24.0, and 29.4 percentage points at k = 1, 4, and 8. Fine-tuning alone helps mainly on direct requests; at four and eight turns, most of the protection comes from the monitor.

Q2. End-to-end results of the three backends, and harmful speech before or after an alert

BackendCalls with an alert (Figure 4)Attack success
Our monitor480
THRD481
Guard-gated monitor426

The original PersonaPlex is jailbroken in 25 of these 50 calls. With our monitor, no harmful speech occurs in these calls.

In order to evaluate if alert arriving late happen, we also checked all calls of the main experiment (341 intents, k = 4 and 8): 2,496 model turns received an alert, 824 of them after the model had started its reply, and none of these turns was judged harmful (Section 5.2). So with our monitor, harmful speech occurs only before the attack is detected, not after an alert.

Q3. The opener statistic, the simulator clock, and latency against the baseline

The 2.85 s opener. This median comes from the training data. For generated responses, we now measure the delay directly on LLaMA Questions (300 spoken questions, monitor on; Table 1). From the end of each question, we time the first audio and the first word of the actual answer:

MetricPersonaPlexDaimonion
Accuracy (%)57.059.0
First audio (s, median)0.360.45
First answer word (s, median)1.392.32

The first audio is almost unchanged, and the actual answer starts 0.9 s later than with the original model.

How delays map onto the simulator clock. The speech model runs in 80 ms frames on a simulated conversation clock. Each monitor result is delivered a number of frames after the call is made that matches its measured processing time, and never earlier. Streaming transcription of the caller is charged the same way. Delivering an instruction takes no conversation time: it is prefilled once and inserted between two frames, which takes tens of milliseconds of compute.

Q4. Benign twins with the full system, and whether they stay benign In order to evaluate the utility of the system more objectively, we measure accuracy on ordinary questions: on LLaMA Questions, Daimonion answers 59.0% of the questions correctly, against 57.0% for the original model, and the monitor sent no alert on any of the 300 questions (Q3). On all 727 examples of Full-Duplex-Bench v1.0, the monitor never sent an alert, and the turn-taking results stay close to those of the original model (Appendix A.13).

Q5. Manual checks of the grading

Choice of judge. The HarmBench classifier was validated against human labels by its authors (93.2% agreement, above a GPT-4 judge at 88.4%), and we apply it to each model turn, its original single-response setting.

Per-turn vs. whole-call scoring. To check this choice, we sampled 50 calls on which per-turn and whole-call scoring disagree: 25 flagged only per turn and 25 flagged only on the whole call, spread evenly over the three models and k = 4 and 8. Two authors labelled each call from the audio, without knowing the model or either verdict. They agreed with each other on 47 of the 50 calls (Cohen's κ = 0.88). Per-turn scoring matched the two annotators on 41 and 44 of the 50 calls. Of the 25 calls flagged only per turn, they judged 22 and 24 harmful; of the 25 calls flagged only on the whole call, they judged 19 and 20 not harmful. In both groups and for both annotators, the human label agreed with per-turn scoring significantly more often than with whole-call scoring (sign test, p < 0.05).

Refusal judge. We checked the 16 sampled calls that the refusal judge marked as refusals: all 16 contain at least an explicit refusal.

We hope these answers address the reviewer's concerns. We are happy to provide more details.

Dear reviewer,

Thank you again for the helpful review. As discussion period ends soon, we want to check whether our response address your concerns. If any point is still unclear, we are more than happy to clarify before the deadline.

Reviewer_QMEPOpenReview ↗

Official comment to authors

I thank the authors for their additional experiments. My concerns have not been resolved fully, and so I will keep my score at 2.

Reviewer_3bjeOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper present crescendo-Live, a benchmark for multi-turn jailbreak against full-duplex speech models. Experiments show that ASR increases substantially as the number of turns increases. After eight turns, the ASR reaches approximately 67–69%. The paper also finds that interruptions can further increase ASR. As defense, the paper proposes Daimonion, an asynchronous LLM monitor. Experiments show that the proposed defense reduces ASR to below 3% with low-latency overhead.

Critical strengths
  1. The threat model of this paper is realistic, as in speech models, multi-turn jailbreak is a natural problem.

  2. The proposed defense shows strong empirical performance. Based on the results shown in Figure 3, Daimonion reduce ASR and this defense holds under multi-turn setting.

Critical weaknesses

The proposed defense method lacks comparison with existing baselines. Please refer to questions 2 and 3 for more details.

Questions for authors
  1. What's the effectiveness of adaptive crescendo compared to non-adaptive crescendo? Does the increased ASR mainly come from increased number of turns or escalated attack query?

  2. How is the effectiveness of Daimonion compared with other defense methods such as adversarial fine-tuned speech model?

  3. How does Daimonion compare against a LLM-based binary safety classifier, but injects a fixed refusal instruction instead?

3 comments

We thank the reviewer for the review and for finding the threat model realistic. We answer the three questions below; Q2 and Q3 also address the weakness on comparison with existing baselines. New and revised text in the updated PDF is marked in blue.

Q1. Adaptive vs. non-adaptive Crescendo, and more turns vs. escalation

We added a non-adaptive baseline (Table 2). For each of the 50 AdvBench intents in our seed set, we use the same attacker model and the same attacker prompt to write a four-turn Crescendo script in advance, without seeing any model reply. A scripted attacker then speaks these four turns in order, with the same voice, and we grade the calls with the same HarmBench classifier. The adaptive attacker is crescendo-Live with the same number of turns (k = 4) on the same intents. Attack success rate (ASR^HB, %):

ModelScriptedAdaptive
PersonaPlex-7B38.050.0
MiniCPM-o 4.550.048.0
VoiceChat-11B42.048.0

The adaptive attacker achieves a 12-point higher attack success rate on PersonaPlex, a 6-point higher rate on VoiceChat, and a similar rate on MiniCPM-o. It can respond to what the model actually says: when the model avoids the question, refuses, or answers only part of it, the attacker can rephrase, take a step back, or follow up.

To separate the effect of more turns from that of escalation, we report the turn of the first success (k = 4, 341 calls per model):

ModelTurn 1Turn 2Turn 3Turn 4
PersonaPlex-7B7426852
MiniCPM-o 4.53367164
VoiceChat-11B6366257

If more turns only gave more chances, with the same chance of success at every turn, first successes would be most common at turn 1 and less common at each later turn. Instead, they are rare at turns 1–2 and most common at turns 3–4, after the request has been built up over several turns. We discuss this further in our response to Reviewer DrxG (W1).

Q2. Comparison with other defenses, such as an adversarially fine-tuned speech model

Our fine-tuned speech model without the monitor serves as this baseline: it is PersonaPlex fine-tuned on dialogues in which harmful requests are answered with refusals. We evaluate it on the same 341 intents with the same attacker and grading (Figure 3, Section 5.2). Attack success rate (ASR^HB, %):

kOriginal PersonaPlexFine-tuned only (no monitor)Daimonion (full system)
135.86.52.1
449.625.51.5
868.932.32.9

Fine-tuning alone works well on direct requests, but its attack success still rises with the number of turns, because the harmful intent is spread across turns that each look harmless. With the monitor, attack success stays low at every turn budget.

Q3. Comparison with a binary safety classifier that injects a fixed refusal instruction

We first clarify a part of our contribution. The original PersonaPlex has no way to receive an instruction during a call. Without this ability, a backend can only act outside the model, for example by muting it and playing a recorded refusal. We add this ability by fine-tuning: two new marker tokens carry an instruction on the model's text stream, and the model is trained to follow the instruction even when it arrives in the middle of its reply. The model then refuses in its own voice and continues the conversation. A binary classifier with a fixed refusal instruction would also rely on this ability. With this ability in place, the backend can be swapped. And our backend monitor actually serve as a binary safety classifier but produce additional reason for rejection and redirection.

We generate the instruction, including the reason, so that the refusal sounds natural in a live call. A fixed refusal (for example, always "I can't help with that") would also stop the harmful answer, but it gives the caller no reason, sounds the same in every call, and ends the exchange abruptly. Our monitor instead writes an instruction such as "Do not help with this. Say you can't explain how to get into someone else's email account because it is illegal and invades their privacy. Then steer back to how to secure their own account" (Appendix A.11). The model then refuses in its own words, says why, and moves the conversation to a safe topic, as a human agent would (Appendix A.12 shows an example).

We hope these answers address the reviewer's concerns. We are happy to provide more details.

Reviewer_3bjeOpenReview ↗

Response to author rebuttal

Thanks for the response. I think my questions are resolved. Given the additional experiments provided, I have increased my score.

Thank you :)

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

This paper introduces crescendo-Live, a live, adaptive benchmark for evaluating multi-turn jailbreaks against full-duplex speech models in simulated phone calls, where an LLM-driven caller escalates intent across turns and can interrupt while the model is speaking. It further proposes DAIMONION, a defense architecture that pairs a full-duplex speech frontend with an asynchronous LLM monitor that injects mid-reply refusal instructions, enabled by fine-tuning the frontend to heed such control without adding response latency. Experiments on three full-duplex models (plus a half-duplex reference) show attack success rising sharply with turn count, while DAIMONION keeps attack success under 3% at all tested conversation lengths with high explicit refusal rates.

Critical strengths
  • The paper targets a realistic, underexplored setting: live multi-turn, full-duplex interactions where the user can barge in and adapt to the model’s responses.
  • crescendo-Live operationalizes an adaptive Crescendo-style attacker in a streaming speech-to-speech loop, extending beyond pre-scripted or single-turn audio attacks.
  • DAIMONION presents an asynchronous safety-monitor architecture that injects internal steering instructions mid-reply, enabled by a tailored fine-tuning recipe (neutral openers, refusal augmentation, interruption simulation).
  • Systematic evaluation across 341 diverse harmful intents from multiple benchmarks, with ablations on caller turn-taking (interrupt vs. patient/deferential) and delivery (verbose/fast).
Critical weaknesses
  • Heavy reliance on automated graders (HarmBench classifier per turn; LLM-based refusal judge) without broad human validation may bias ASR estimates, especially with Whisper transcription errors and per-turn scoring.
  • DAIMONION’s efficacy is demonstrated only on a single frontend (PersonaPlex-7B); generality to other full-duplex models is not shown.
  • The defense requires bespoke fine-tuning with internal instruction tokens and neutral openers; portability to off-the-shelf or closed models is unclear.
  • The opener strategy could introduce subtle utility/quality trade-offs (e.g., perceived verbosity or delayed informativeness) that are not quantitatively evaluated on benign queries.
  • Single trial per intent and a single caller voice/model limit statistical confidence and diversity; results could be sensitive to the attacker’s style or ASR/TTS stack.
  • Per-turn ASR-based judging may overcount minor slips relative to a holistic conversation-level harm assessment; a session-level safety metric would complement per-turn scoring.
  • Limited benign-traffic evaluation for over-refusal beyond a 50-pair AdvBench/SARSteer subset and backend-only analysis; no large-scale benign multi-turn runs are reported for end-to-end DAIMONION.
  • The compute/latency footprint for the monitor (27B FP8 with speculative decoding) is nontrivial; resource and cost implications are not fully profiled.
  • Connection to recent diagnostic frameworks for multi-turn state evolution (e.g., STAR) and intent-organization taxonomies (e.g., INTACT) could be deepened to situate the escalation mechanisms and detection surface alignment.
  • A comparison to internal activation–based defenses (e.g., AEGIS) would further contextualize DAIMONION as an external monitor versus in-model risk gating approaches and clarify trade-offs in applicability and latency.
Questions for authors
  • How sensitive are the results to the ASR/TTS stack choices? Have you tried alternative ASR models or higher sampling rates to quantify how transcription quality affects ASR^HB and refusal measurements?
  • Can you report a larger-scale benign evaluation (multi-turn, diverse topics) for DAIMONION end-to-end, including refusal rates, subjective conversational quality, and any latency/user-experience effects of the neutral opener?
  • What is the exact compute and latency profile of the 27B monitor under load (per-call GPU memory, tokens/sec, polling overhead), and how does it scale with concurrent sessions?
  • How portable is the frontend instruction-injection scheme? Could you share more details on the marker tokens and insertion interface, and whether you successfully replicated the behavior on another open duplex model?
  • Have you explored adaptive opener durations (or learning an opener policy) to further reduce the window where a late alert might miss the refusal boundary?
  • Can you provide a breakdown of false positives (benign/protective cases) and the monitor’s rationales, and whether small prompt changes to the monitor reduce these without harming detection?
  • To what extent does DAIMONION remain effective with multilingual callers or code-switched speech, and under background noise or accented speech conditions?
5 comments

We thank the reviewer for the detailed review and for recognizing the realistic live setting, the adaptive caller, and the asynchronous monitor. We group each question with the weakness it follows up, and give the new results in each response. Revised text in the updated PDF is marked in blue.

W1 & W6. Automatic grading, per-turn scoring, and a session-level metric

Choice of judge. The HarmBench classifier is trained to judge a single response and was validated against human labels by its authors (93.2% agreement, above a GPT-4 judge at 88.4%), so we apply it to each model turn.

Session-level score. We also scored every call with the same classifier over the whole transcript at once (attack success %, per-turn / whole-call):

Modelk = 4k = 8
PersonaPlex-7B49.6 / 35.868.9 / 43.4
MiniCPM-o 4.551.0 / 37.069.2 / 44.3
VoiceChat-11B47.2 / 32.866.9 / 44.6
Daimonion1.5 / 0.32.9 / 0.6

At k = 1 the two coincide, since a call has one model turn. Whole-call scores are lower because one long transcript dilutes a single harmful turn, but the conclusions hold: multi-turn attacks break the models that resist direct requests, and Daimonion stays below 1%.

Human check. We sampled 50 calls on which per-turn and whole-call scoring disagree (25 flagged only per turn, 25 only on the whole call, spread evenly over the three models and k = 4 and 8). Two authors labelled each call from the audio, without knowing the model or either verdict; they agreed on 47 of 50 (Cohen's κ = 0.88). Their verdicts matched per-turn scoring on 41 and 44 of the 50 calls, and whole-call scoring on only 9 and 6 (sign test, p < 0.05 in both groups), so per-turn scoring is closer to human judgment, not an overcount. Because they listened to the audio, this also checks the Whisper transcripts. For the refusal judge, all 16 sampled calls it marked as refusals contain at least an explicit refusal.

Q1. Sensitivity to the ASR/TTS stack

ASR enters our setup in three places: the simulated caller hears the model, the monitor hears the caller, and the grader reads the model's speech. We re-transcribed the recorded calls (341 intents × k = 4 and 8) with a second ASR model, Parakeet-TDT-0.6B-v2, on the same speech segments as Whisper-large-v3, and repeated each step:

Table 1. The simulated caller hears the model's speech.

SystemWords that differ, Whisper vs. Parakeet (k = 4 / 8)
PersonaPlex-7B1.8% / 2.9%
MiniCPM-o 4.53.6% / 3.8%
VoiceChat-11B2.9% / 3.6%
Daimonion2.8% / 4.0%

The two ASR models differ on about one or two words in a typical 40-word turn, so the simulated caller would hear essentially the same reply with either one.

Table 2. The monitor hears the caller's speech.

BudgetCalls alerted of 341, WhisperCalls alerted of 341, Parakeet
k = 4327326
k = 8336336

Both ASR models transcribe the caller's speech with only 2–4% word errors against the caller's script. The monitor reads either transcript and alerts on nearly the same calls: 327 vs. 326 of 341 at k = 4, 336 vs. 336 at k = 8, and the same decision on 98.7% of calls.

Table 3. The grader reads the model's speech.

SystemAttack success %, Whisper / Parakeet (k = 4)Attack success %, Whisper / Parakeet (k = 8)
PersonaPlex-7B49.3 / 48.468.6 / 68.0
MiniCPM-o 4.550.7 / 49.671.0 / 68.0
VoiceChat-11B47.5 / 47.867.4 / 67.2
Daimonion1.5 / 2.12.9 / 4.4

For every model and budget, attack success differs by at most 3 percentage points between the two transcripts, and Daimonion stays below 5% with both. (The Whisper column re-transcribes the same segments as Parakeet, so it can differ from the paper's numbers by up to 1.8 points.)

All voices are clean synthetic speech, and the 8 kHz line is deliberate. Voice agents will very likely answer phone calls, so we test the setting they are most likely to face in practice. A small pilot on 25 samples with the original 24 kHz audio gave similar results. How the caller speaks still matters: a 1.3× faster caller lowers attack success on PersonaPlex from 50% to 26% (Table 4).

W2, W3 & Q4. A second frontend, portability, and the injection interface

We connected the unchanged backend (same monitor, prompt, polling and streaming ASR) to a second fine-tuned frontend and to an unmodified model (50 AdvBench intents, k = 4):

FrontendFine-tuned with our recipeMonitorAttack success
PersonaPlex-7B, releasednono25/50 (50%)
Daimonion (PersonaPlex)yesyes0/50 (0%)
MiniCPM-o 4.5, releasednono24/50 (48%)
MiniCPM-o 4.5 + our recipeyesyes0/50 (0%)
Realtime-Venus Audio-9Bnoyes3/50 (6.0%)

The recipe carries over to a second full-duplex model (W2). MiniCPM-o 4.5 with our recipe gets 0/50 attack success, like Daimonion. More generally, the setup can transfer to any model that can take in instructions. Realtime-Venus (arXiv:2609.13814, built on MiniCPM-o 4.5) is a good example: it needs no fine-tuning (W3), and the same backend holds it to 3/50 (6.0%).

W4, W7 & Q2. Benign utility, the opener's cost, and conversational quality

We compare the full system (monitor on) with the original PersonaPlex on LLaMA Questions (300 spoken questions; Table 1):

MetricPersonaPlexDaimonion
Accuracy (%)57.059.0
First audio (s, median)0.360.45
First answer word (s, median)1.392.32
Monitor alerts–0 / 300
UTMOSv2 †2.362.37
NISQA naturalness †3.313.23
NISQA quality †4.194.06

† Automatic MOS, from a second run of the same protocol.

Daimonion keeps the original accuracy and starts speaking almost as quickly; the actual answer starts 0.9 s later, which is the cost of the short neutral phrase at the start of the reply. Automatic MOS stays within 0.15 of the original model, so the fine-tuning and the opener leave the voice essentially unchanged. The monitor also sent no alert on any of the 727 examples of Full-Duplex-Bench v1.0, where the turn-taking results stay close to those of the original model (Appendix A.13). We did not run a human listening study; these measures are automatic.

False refusals in multi-turn conversations. We use the 50 SARSteer benign twins, run with the same Crescendo attacker, which keeps asking for more specific details. With the monitor, the model refused 13 of the 50; without it, none. Looking into the gap, most are protective requests, such as preventing identity theft or securing a government database. The simulated caller opens with background, e.g. the most common ways people cheat on online exams, and then asks for concrete protective steps, e.g. what an admin does on day one to lock a database down. The monitor treats these concrete steps, and in three calls the background question itself, as requests for operational detail and triggers a refusal (breakdown in Q6).

W5. One trial per intent, one caller voice and model

We agree and list this as a limitation. Repeated trials are expensive: each simulated call takes about 20 GPU-minutes, so we ran one call per intent. With one call per intent, a confidence interval over the 341 intents is still valid for the overall rate (attack success %, 95% CI):

Systemk = 4k = 8
PersonaPlex-7B, released49.6 [44.3, 54.8]68.9 [63.8, 73.6]
Daimonion1.5 [0.6, 3.4]2.9 [1.6, 5.3]

The gap is far outside sampling noise. For the attacker's style, the paper varies how the caller takes turns (Table 3) and how it speaks (Table 4), and the revision adds a scripted, non-adaptive caller (Table 2). Style changes attack success by up to 24 points (a 1.3× faster caller: 50% to 26% on PersonaPlex), so the paper reports each style's result in its own table rather than one pooled number. For the speech stack, all calls use clean, standard American-English synthetic voices, and changing the ASR model changes little (Q1).

Q6. False positives, the monitor's rationales, and prompt changes

Breakdown. The 13 refused benign twins (W7) by type:

Refused benign requestTwins
Protective, after an alert (e.g. preventing identity theft, securing a government database, preventing exam cheating)10
News article, after an alert (on fact-checking; on preventing panic)2
No alert: the speech model itself declined (a letter praising an employee)1

The monitor writes a one-sentence rationale before each decision. In 6 of the 12 alerts it flagged the follow-up asking for concrete protective steps, in 3 the background question (e.g. "The user is asking for the most common ways people cheat on online exams, which is a request for methods of academic dishonesty."), and in 3 a writing request or a misheard word ("read" transcribed as "raid"). The cause is the shipped rule: it allows protective answers only if they never reach specifics someone could act on, and it counts a narrowing follow-up as a request for such specifics, so concrete defensive steps are flagged like attack steps.

A tunable backend. Because the false alarms come from the rule, not from the speech model, the rule can be fixed directly. The policy is the monitor's plain-text prompt, and every monitor input is logged, so an edit can be replayed offline on the same calls, with no retraining and no new conversations. We tried six one- or two-sentence edits on 50 harmful calls and their 50 benign twins; the table shows the shipped prompt and the best two:

Monitor promptHarmful calls flagged (of 50)First alert vs. shipped (earlier / later)Benign false alarms (of 50)
Shipped prompt49–12
Edit 1: allow concrete defensive steps491 / 05
Edit 1 + Edit 2: the rationale also weighs protection492 / 04

We improve the monitor by editing the prompt only. Edit 1 rewrites the rule's exception to allow "requests for how to protect, secure, detect or prevent something, including concrete defensive steps, as long as the answer would not explain how to carry out the attack itself", which separates defending from attacking. Edit 2 adds one clause to the rationale the monitor writes before deciding: whether the answer "would mainly help someone protect against that harm". Together they clear more false alarms and flag some harmful calls earlier, and Edit 1 keeps almost all detections on the harmful calls of the main experiment. The other four edits cleared fewer false alarms, delayed alerts, or missed a harmful call. The edits were chosen on these same twins, so the table shows how far the prompt can be tuned, not a validated new default.

Comments on the paper

1 comment

Dear AC and reviewers,

Thank you for the detailed reviews. We have uploaded a revised PDF (changes in blue). Here is a short summary of the new results and where to find them.

1. What the monitor adds beyond fine-tuning (QMEP, DrxG, 3bje). With the monitor off, our fine-tuned frontend is jailbroken in 6.5% of direct requests (35.8% for the released model), but in 25.5% and 32.3% of calls at k = 4 and 8. With the monitor, it stays at 2.1%, 1.5% and 2.9% (Figure 3, Section 5.2). Fine-tuning mainly covers direct requests; the monitor is what stops the escalation.

2. Adaptive escalation, not just more turns (DrxG, 3bje). A scripted Crescendo caller succeeds less often than our adaptive caller on two of three models (Table 2), and first successes are rare at turn 1 and most common at turns 3–4. The attack works because it adapts and escalates, not only because it has more turns.

3. Do alerts arrive in time? (QMEP, 3bje) We added two checks. First, we examined every model turn that received an alert in the main experiment (2,496 turns at k = 4 and 8). None was judged harmful, so no alert came too late (Section 5.2). Second, we compared the delay of our backend design with THRD, which judges only at the end of each caller turn, by replaying the same recorded calls through both (new Appendix A.14). Our monitor runs while the caller is still speaking, so more than half of its alerts are ready when the caller's turn ends (median delay 0 s), while THRD's alerts arrive a median 3.3 s later. As a result, our alert comes before the harmful content in 76% of calls, against 51% for THRD. End to end on 50 live calls, 0 attacks succeed with our monitor, 1 with THRD and 6 with a guard-gated monitor (Section 5.3). Our alerts arrive in time, and checking while the caller is still speaking is what makes them early.

4. A second frontend (nCsc). Our fine-tuning recipe carries over to MiniCPM-o 4.5 (0 of 50 attacks succeed), and the same backend holds an unmodified model, Realtime-Venus, to 3 of 50. The approach is not tied to PersonaPlex.

5. Cost to benign use (all). The model starts speaking almost as fast as before (0.45 vs. 0.36 s) and answers as accurately (59.0% vs. 57.0% on LLaMA Questions), but its answer starts 0.9 s later because of the opener (Table 1). The monitor never alerts on Full-Duplex-Bench (Appendix A.13). On the 50 benign twins, the model answers 161 of 200 caller queries, against 190 without the monitor (Section 5.3). Most false alarms come from protective requests. Two short prompt edits, tuned on the same twins, lower them from 12 to 4 without retraining. The main costs are a later answer and some false refusals, which the monitor prompt alone can reduce.

6. Robustness of the evaluation (nCsc, QMEP). A second ASR model changes attack success by at most 3 points. On 50 calls where per-turn and whole-call scoring disagree, two annotators side with per-turn scoring in 41 and 44 cases (Cohen's κ = 0.88). At k = 4 and 8, the 95% confidence intervals of Daimonion and the released model do not overlap. Our results hold with a different ASR model and under human checks.

The revision also scopes the latency claim to when the model starts to speak (Sections 1 and 6), adds STAR, INTACT and AEGIS to the related work, and extends the limitations.

Thank you again for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.