acceptodds
Under review as a conference paper at ICLR 2027

One Turn Too Late: Learning When to Intervene Against Multi-Turn Malicious Intent

Abstract

Hidden malicious intent in multi-turn dialogue poses a growing threat to deployed large language models (LLMs). Rather than exposing a harmful objective in a single prompt, attackers can distribute their intent across multiple benign-looking turns, making defense a problem not only of whether a dialogue is harmful, but also of when intervention becomes necessary. Existing trace-level labeling approaches provide only coarse safety signals and do not identify this intervention boundary, making it difficult to distinguish timely intervention from premature refusal or a block that comes too late. This work introduces turn-level harm-enabling supervision for multi-turn defense. We define the earliest harm-enabling turn as the first point at which delivering a candidate response would make the accumulated interaction sufficient to enable harmful action. To instantiate this supervision at scale, we construct the Multi-Turn Intent Dataset (MTID), which contains adaptive attack rollouts, matched benign hard negatives, and annotations of this boundary. Using MTID, we train TurnGate, a response-aware monitor that learns when to intervene, and further optimize its policy through multi-turn reinforcement learning. Experiments show that turn-level boundary supervision improves intervention localization, while reinforcement learning further improves the safety–utility trade-off. TurnGate outperforms existing guardrails and multi-turn monitoring baselines, and generalizes across risk domains, attacker pipelines, and target models. Our code is available at https://anonymous.4open.science/r/TurnGate-ICLR.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Safety, Defense, safety--utility trade-off

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_ST9zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper constructs MTID with automatically annotated intervention boundaries and trains a response-aware monitor through weighted supervised fine-tuning and offline reinforcement learning.

Critical strengths
  • Intervention timing is a relevant problem, and considering candidate responses alongside dialogue history is well motivated.

  • Matched benign examples, seed-level splits, and adaptive online evaluations are useful experimental choices.

Critical weaknesses
  1. Limited novelty. The contribution largely combines existing attack generation, automated judging, supervised learning, and policy optimization. Its engineering emphasis and limited insights weaken its suitability for ICLR.
  2. Limited attacker diversity. Training and primary evaluations rely on CKA. Beyond MAJ, independent multi-turn attacks, including ActorAttack, are needed to assess robustness beyond CKA-specific patterns.
  3. Missing targeted baseline. Since the annotation pipeline evaluates sufficiency given a known harmful goal, a natural deployment baseline would first infer plausible harmful goals from the dialogue history, then assess whether the accumulated dialogue and candidate response provide sufficient information to achieve them. The paper does not clearly evaluate this explicit two-stage alternative.
Questions for authors

How does TurnGate compare with an explicit goal-inference-then-sufficiency monitor under identical information constraints?

5 comments
AuthorsOpenReview ↗

Response to Reviewer ST9z (Part 1)

We thank Reviewer ST9z for the insightful review and for finding the response-aware design well motivated. Below, we respond to the questions on novelty, attacker diversity, and the goal-inference-then-sufficiency baseline.

W1. Novelty and insight

To our knowledge, TurnGate is the first work to formulate multi-turn defense as an information-sufficiency boundary localization problem. We provide explicit supervision for this boundary and optimize intervention timing. The intervention target is the first turn at which releasing a candidate response makes the accumulated dialogue sufficient to realize a harmful objective.

Identifying harmful intent alone does not tell us when to intervene. Harmful intent may be present from the beginning of a dialogue, while the information needed to realize it accumulates across turns. Blocking too early sacrifices dialogue utility. Blocking too late allows sufficient harmful information to be delivered. With MTID and a sequential stopping objective, we make this safety–utility trade-off an explicit learning problem.

Table 1 shows why intervention timing matters. Trajectory-level Naive-SFT has few misses (8.9%), yet it blocks early on 62.9% of harmful trajectories and preserves only 60.9% of benign sessions. Uniform turn-level SFT preserves 93.0% of benign sessions but misses 73.8% of harmful trajectories. TurnGate achieves the highest exact-turn accuracy (41.4%) and overall score (0.699).

W2. Independent attacker diversity

Thank you for suggesting ActorAttack [8]. We added an evaluation on 50 objectives to further assess TurnGate's robustness beyond CKA, complementing the existing independent MAJ [9] evaluation in Fig. 4d.

For MAJ, ASR falls from 0.780 to 0.631 for Chemistry →\rightarrow Cybersecurity and from 0.715 to 0.593 in the reverse direction.

We evaluate ActorAttack on 50 objectives fixed before observing outcomes, with shared initial plans, a GPT-5.2 target, three actors per objective, up to three rewrites after refusals, a Qwen3-8B-abliterated controller, and Gemini-2.5-Flash judging. Candidates are gated before delivery, and follow-ups depend on delivered responses.

MonitorSuccessful attacks / 50 ↓\downarrow
No defense17/50
Goal →\rightarrow sufficiency, GPT-5.211/50
Reweighted-SFT6/50
TurnGate5/50

TurnGate reduces ActorAttack ASR from 34% to 10%, a relative reduction of approximately 71%. Together with the consistent improvements against MAJ, these results provide evidence that TurnGate generalizes beyond the CKA attack patterns used in training.

CKA remains our primary stress test because The Trojan Knowledge [1] reports higher HarmBench Full-Success rates than MAJ and ActorBreaker on GPT-oss-120B and Claude-Haiku-4.5. ActorAttack [8] refers to the v1 paper. The prior comparison uses the name ActorBreaker.

AuthorsOpenReview ↗

Response to Reviewer ST9z (Part 2)

W3 & Q1. Goal inference followed by sufficiency evaluation

Thank you for proposing this baseline. The judge uses the ground-truth harmful objective during annotation and evaluation. TurnGate has no access to it at inference time and sees only the dialogue history, current query, and candidate response. Under the same information constraints, TurnGate achieves an overall score of 0.667, compared with 0.162 and 0.154 for the two-stage Qwen3-4B and GPT-5.2 baselines.

In the first stage, the monitor infers up to three plausible harmful objectives from the observable dialogue. In the second stage, it checks whether the accumulated dialogue, including the candidate response, contains sufficient information to achieve each objective. For example, given a series of individually benign cybersecurity questions, the first stage may infer an unauthorized-access objective. The second stage then checks whether the information accumulated so far is sufficient for unauthorized access.

We use the same A/B/C/D sufficiency rubric as in our annotation pipeline. The monitor outputs BLOCK if any inferred objective is judged as Full Success (D). The implementation also triggers BLOCK upon an explicit provider refusal. Intention Analysis [7] assesses inferred intent. This baseline evaluates the sufficiency of the accumulated information for the inferred objective.

Supplementary evaluationIntercepted by boundary ↑\uparrowExact-boundary block ↑\uparrowBenign all-pass ↑\uparrowOverall ϕ2\phi_2 ↑\uparrow
Goal →\rightarrow sufficiency, Qwen3-4B3/34 (8.8%)3/34 (8.8%)59/60 (98.3%)0.162
Goal →\rightarrow sufficiency, GPT-5.2, medium reasoning4/34 (11.8%)2/34 (5.9%)37/60 (61.7%)0.154
TurnGate22/34 (64.7%)15/34 (44.1%)53/60 (88.3%)0.667

The evaluation panel contains 34 Full-Success harmful traces and 60 benign traces selected by an outcome-blind per-seed rule. Both stages receive the complete observable context. Qwen3-4B-Instruct-2507 uses greedy decoding and 512 output tokens per stage. GPT-5.2 uses medium reasoning and up to 4,096 tokens per stage, including reasoning. We report each baseline's highest overall score among ordinary-goal and capability-oriented prompt variants.

TurnGate intercepts more harmful trajectories than either two-stage baseline while preserving 88.3% of benign dialogues. Even with GPT-5.2, explicitly prompting the model to infer goals and assess sufficiency performs substantially worse than TurnGate.

References

[1] Wei et al. The Trojan Knowledge: Bypassing Commercial LLM Guardrails via Harmless Prompt Weaving and Adaptive Tree Search. arXiv preprint, 2025.

[7] Zhang et al. Intention Analysis Makes LLMs A Good Jailbreak Defender. COLING, 2025.

[8] Ren et al. Derail Yourself: Multi-turn LLM Jailbreak Attack through Self-discovered Clues. arXiv preprint, 2024, v1.

[9] Srivastav and Zhang. Safe in Isolation, Dangerous Together: Agent-Driven Multi-Turn Decomposition Jailbreaks on LLMs. REALM Workshop, 2025.

We thank the reviewer for recognizing the motivation behind our response-aware design and the value of our experimental evaluations. We appreciate the valuable feedback and would be happy to provide further clarification if any concerns remain.

Reviewer_JAv1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the problem of turn level supervision for multi-turn defense for LLMs. They train TurnGate which is a response-aware monitor that learns when it is a good idea to intervene. As a side product, they also provide the multi-turn intern dataset, which contains attack that are posed in multiple turns. This work is in contrast with standard safeguards that assess individual each turn individually or each conversation as whole.

Critical strengths

LLMs are used in multi-turn settings, and as such, they are susceptible of multi-turn attacks. Tackling this problem and providing more advanced safeguards and a dataset for further development is definitely a good contribution that may merit publication.

The authors also study benchmark shift (e.g., train on chemistry but evaluate on cybersecurity), target-model shift, and attacker pipeline shift (evaluate on conversations produce by a different attacker pipeline). This strengthens the paper findings and showcases the approach. The paper also evaluates multiple models from GPT and Qwen family

Critical weaknesses

There is some recent or not that recent work that should have probably been cited https://arxiv.org/abs/2606.01738, https://arxiv.org/pdf/2510.15017, https://arxiv.org/pdf/2503.00187

It would have been good to update the models being evaluated in the paper, as they are 1-2 generations behind. LLMs are improving rapidly and by the time the conference starts, the learnings in the paper may change just because of that.

Questions for authors
  • this paper may merit a mention, https://arxiv.org/pdf/2505.06120. I know it is not on the same problem, but the fundamental issue of LLMs getting lost and failing to comply with what they have been trained on over multiple turns is essentially core of your paper and this paper.

The improvements over Intention analysis are not that large. Would it be possible to provide error bars in Figure 3 and the results in the table so you demonstrate whether there are statistical significance results?

4 comments
AuthorsOpenReview ↗

Response to Reviewer JAv1

We thank Reviewer JAv1 for the insightful review and for recognizing our evaluations under benchmark, target-model, and attacker-pipeline shifts. Below, we respond to the questions on related work, newer models, and uncertainty estimates.

W1 & Q1. Related work

Thank you for these references. We will add them to Sec. 2 and distinguish their monitoring and control approaches from our boundary-supervision objective:

WorkConnection and distinction
THRD [3]Aggregates temporal risk without training; we learn intervention timing from response-conditioned sufficiency labels.
Active Honeypot Guardrail System [4]Probes intent through interaction; TurnGate decides whether to deliver an existing candidate response.
Steering Dialogue Dynamics [5]Studies sequential safety control; we annotate and learn the first information-sufficiency boundary.
LLMs Get Lost [6]Documents degradation over multiple turns, supporting the motivation for evaluating safety across a dialogue.

Our contribution is the explicit intervention target, MTID's boundary supervision, and the joint evaluation of timing and benign preservation.

W2. Evaluated model versions

Thank you for suggesting evaluation on newer models. We added target-model evaluations on GPT-6 Luna and Claude Haiku 5.5 using a subset of 30 tasks, with ASR measured as the fraction of tasks on which the attack succeeds. TurnGate reduces ASR from 26/30 (86.7%) to 19/30 (63.3%) on GPT-6 Luna and from 23/30 (76.7%) to 18/30 (60.0%) on Claude Haiku 5.5, decreases of 23.3 and 16.7 percentage points, respectively, computed from the unrounded rates. We report these alongside the existing Gemini target result from the paper:

TargetNo defense ASR ↓\downarrowTurnGate ASR ↓\downarrow
Gemini-3.1-Pro (paper, Fig. 4)85.3%67.6%
GPT-6 Luna (added)26/30 (86.7%)19/30 (63.3%)
Claude Haiku 5.5 (added)23/30 (76.7%)18/30 (60.0%)

These additional evaluations extend the target-model coverage beyond the original experiments. The paper also evaluates GPT-5.2 and GPT-OSS-120B as prompt-based monitors (Table 1).

Qwen3-4B is the trained monitor, shared with trainable baselines to compare training methods on the same backbone. Appendix E explains this choice in terms of compute and per-turn inference cost. The target-transfer results test the monitor when a different assistant generates the responses.

Q2. Improvement over Intention Analysis and uncertainty

Thank you for requesting uncertainty estimates. Table 1 reports an overall ϕ2\phi_2 improvement from 0.42 to 0.70 (+0.28) over GPT-5.2 Intention Analysis [7], with cross-domain gains of +0.108 and +0.097 in Table 2.

We additionally report sensitivity intervals for the differences below:

MetricGPT-5.2 Intention AnalysisTurnGateDifference: 95% interval
Intercepted by annotated boundary ↑\uparrow38.33%82.50%+44.17 pp [24.62, 63.71]
Exact-boundary block ↑\uparrow20.67%41.08%+20.42 pp [1.22, 39.62]
Benign all-pass ↑\uparrow74.58%83.42%+8.83 pp [−7.41, 25.08]
Overall ϕ2\phi_2 ↑\uparrow0.420.70+0.28 [0.076, 0.494]

These calculations condition on a TurnGate benign pass count of 1001/1200 and use 60 effective seeds per class, conservative bounds for unknown cross-method pairing, and a delta-method approximation for the overall score.

References

[3] Ma et al. THRD: A Training-Free Multi-Turn Defense Framework for Jailbreak Attacks on Large Language Models. arXiv preprint, 2026.

[4] Wu et al. Active Honeypot Guardrail System: Probing and Confirming Multi-Turn LLM Jailbreaks. arXiv preprint, 2025.

[5] Hu et al. Steering Dialogue Dynamics for Robustness against Multi-turn Jailbreaking Attacks. arXiv preprint, 2025.

[6] Laban et al. LLMs Get Lost In Multi-Turn Conversation. arXiv preprint, 2025.

[7] Zhang et al. Intention Analysis Makes LLMs A Good Jailbreak Defender. COLING, 2025.

We thank the reviewer for the valuable comments and positive feedback in recognizing our paper's major contribution. We are grateful to answer any follow-up questions.

Reviewer_JAv1OpenReview ↗

Authors addressed my questions

I will maintain my score but i am happy to increase it during the discussion period.

Thank you for your positive feedback and willingness to further support our work. We sincerely appreciate your time and consideration!

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

This paper studies safety defense in multi-turn conversations, where attackers may gradually obtain harmful information through a sequence of seemingly benign questions. The authors define the earliest turn at which the accumulated information becomes sufficient to achieve a harmful objective as the intervention point. They propose TurnGate, a defense model that examines the conversation history and the candidate response to decide whether to block it. The model is trained using supervised fine-tuning followed by reinforcement learning.

The authors construct the MTID, containing 8,000 harmful and 8,000 benign conversations, with intervention points annotated using automated evaluators. They evaluate TurnGate through offline experiments, online tests against adaptive attackers, and generalization tests across risk domains, target models, and attack methods. The results show that TurnGate outperforms existing defenses on the proposed evaluation metrics and reduces attack success rates in online tests.

Critical strengths

The paper addresses an important and practical problem: when to intervene against harmful information in multi-turn conversations. The authors construct a dataset and evaluate their method against adaptive attackers in online experiments, demonstrating reduced attack success rates. These results provide some empirical evidence of the method's potential practical value.

Critical weaknesses

The authors define the earliest turn at which the accumulated information becomes sufficient to achieve a harmful objective as the correct intervention point, but do not adequately justify why this is the appropriate time to intervene. Since the dataset labels, model training, and evaluation metrics all rely on this definition, improvements in experimental results do not necessarily demonstrate that the model has learned when intervention should actually occur. This fundamental issue limits the practical significance of the paper's findings.

Questions for authors

Could the authors explain why the earliest turn at which sufficient information has been accumulated to achieve a harmful objective t* should be considered the correct intervention point? Does this definition still hold if earlier responses have already increased the risk of harm, or if the user already possesses relevant knowledge? This question directly affects whether the dataset labels, model training, and evaluation metrics truly reflect appropriate safety intervention decisions.

3 comments
AuthorsOpenReview ↗

Response to Reviewer dhDT

We thank Reviewer dhDT for the insightful review and for recognizing the practical importance of deciding when to intervene in multi-turn conversations. Below, we respond to the question of why t∗t^{\ast} is the intervention target.

W1 & Q1. Why this intervention boundary?

Thank you for raising this question. Earlier responses may increase the risk of harm. An attacker's prior knowledge may also make otherwise insufficient information actionable. Blocking any information that could contribute to harm would lead to substantial over-refusal. For example, knowledge about chemical properties or reaction mechanisms can support legitimate scientific work as well as harmful objectives. A practical defense needs to preserve legitimate dialogue while limiting harmful information disclosure.

We focus on whether the information delivered by the AI system becomes sufficient to realize a harmful objective. Under this objective, t∗t^* is the latest point at which the system can prevent the first sufficient prefix from being delivered. Earlier intervention interrupts information exchange before that boundary. Later intervention would allow the sufficient prefix to be delivered. TurnGate checks each candidate response before delivery and can intervene at this boundary.

An attacker's private knowledge can move the effective boundary earlier, as discussed in Appendix F.1. The annotated t∗t^* estimates the intervention target under our information-sufficiency criterion. Its interpretation depends on the information available in the dialogue and the sufficiency assessment.

Appendix C.2 evaluates cumulative dialogue prefixes. Appendix C.6 independently checks sufficiency at t∗t^* and insufficiency at t∗−1t^*-1. To evaluate the practical effects of the learned intervention policy, Figures 3–4 measure ASR in closed-loop adaptive interactions. Tables 1–2 measure benign preservation.

Reviewer_dhDTOpenReview ↗

Official Comment by Reviewer dhDT

Thank you for the clarification. The formulation and Appendices C.6–C.7 clarify the operational meaning of t, while F.1 acknowledges its dependence on prior knowledge and judge choice. I still regard it as a task-specific boundary rather than a universally optimal intervention point, but I consider my main concern largely addressed. I already raised my rating from 2 to 3.

We thank the reviewer for the valuable comments and positive feedback in recognizing our paper's major contribution. We would be happy to answer any follow-up questions.

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

The paper introduces MTID, a multi-turn dialogue dataset annotated with the first harm-enabling turn, and TURNGATE, a response-aware monitor that decides whether to release each candidate response. TURNGATE is trained through turn-level supervised fine-tuning followed by offline reinforcement learning.

Critical strengths

1.The paper addresses an important problem: deciding when to intervene in multi-turn interactions while avoiding unnecessary refusals.

2.MTID includes matched benign hard negatives, allowing the evaluation to consider both harmful information release and over-refusal.

3.The experiments include closed-loop adaptive attacks and additional evaluations across different domains, target models, and attacker pipelines.

Critical weaknesses
  1. I am not convinced that the experiments clearly establish where the improvement comes from. In Table 1, turn-level Naive-SFT performs worse than trajectory-level Naive-SFT, so the claim that turn-level supervision is the key improvement needs stronger support.

  2. To me, the main contribution is the turn-level labeling scheme, while the need for the additional RL stage is less convincing. Since the online experiments omit Reweighted-SFT, it is difficult to see what this extra stage actually adds against adaptive attackers.

  3. I find the robustness claims too optimistic, given that the attack success rate reaches greater than 0.60 at the largest tested budget in Fig 3. This is an improvement over the baselines, but the remaining failure rate deserves more attention when discussing the method’s practical value.

Questions for authors
  1. Why are all turns before t* labeled PASS, even though some responses may already provide useful harmful information? How does this labeling account for attackers who can combine these responses with prior knowledge?

2.Could the authors report benign pass-through rates for the online out-of-distribution tests in Fig 4? This would help distinguish better safety decisions from simply more conservative blocking.

4 comments
AuthorsOpenReview ↗

Response to Reviewer DffK (Part 1)

We thank Reviewer DffK for the insightful review and for recognizing the importance of deciding when to intervene while avoiding unnecessary refusals. Below, we respond to the questions on the source of the improvement, the value of RL, robustness at larger budgets, pre-boundary labels, and online benign pass-through.

W1. What improvement comes from turn-level supervision?

Thank you for raising this point. Trajectory-level supervision does not explicitly tell the model when to intervene. This helps explain the high rate of early blocking. Turn-level supervision provides this timing signal. Many turns are labeled PASS before the critical BLOCK turn, creating a label imbalance. Uniform SFT tends to favor PASS. It preserves 93.0% of benign dialogues but misses 73.8% of harmful trajectories.

Training methodMiss ↓\downarrowExact-turn block ↑\uparrowBenign pass-through ↑\uparrowOverall ϕ2\phi_2 ↑\uparrow
Trajectory-level Naive-SFT8.9%28.2%60.9%0.568
Turn-level Naive-SFT73.8%16.3%93.0%0.361
Reweighted-SFT37.9%34.3%84.0%0.610
TurnGate17.7%41.4%83.4%0.699

Reweighting PASS and BLOCK reduces misses from 73.8% to 37.9% and improves exact-turn accuracy from 16.3% to 34.3%. Benign preservation remains higher than with trajectory-level SFT (84.0% vs. 60.9%). RL further reduces misses to 17.7% and improves exact-turn accuracy to 41.4%, with little change in benign preservation.

W2. Incremental value of reinforcement learning

Thank you for suggesting the online Reweighted-SFT control. RL lowers ASR by 23.5 and 11.8 percentage points at search budgets 3 and 5, respectively.

Reweighted-SFT already improves substantially over uniform turn-level SFT, raising the overall score from 0.361 to 0.610 (Table 1). SFT learns individual PASS/BLOCK decisions. The quality of the defense depends on when the first BLOCK occurs across the dialogue. RL directly optimizes this stopping objective through action-dependent process rewards, accounting for the consequences of blocking too early or missing the harmful boundary.

The offline ablation shows that RL reduces harmful misses from 37.9% to 17.7% and raises exact-turn accuracy from 34.3% to 41.4%. Benign pass-through changes only slightly from 84.0% to 83.4% (Table 1). The overall score improves from 0.610 to 0.699. The cross-domain scores improve from 0.352 to 0.543 and from 0.390 to 0.578 (Table 2).

We added the requested online comparison on the same 17 objectives, using GPT-5.2, a shared Qwen3-8B-abliterated CKA controller [1], Gemini-2.5-Flash judging, and matched search depth and branching. Every candidate is gated before attacker scoring and observation:

Search budgetNo defense ASR ↓\downarrowReweighted-SFT ASR ↓\downarrowTurnGate ASR ↓\downarrow
315/17 (88.2%)13/17 (76.5%)9/17 (52.9%)
517/17 (100%)12/17 (70.6%)10/17 (58.8%)

Each budget uses a fresh adaptive search on the same 17 objectives.

AuthorsOpenReview ↗

Response to Reviewer DffK (Part 2)

W3. Residual attack success at larger budgets

We agree that TurnGate has room for improvement at the largest attack budget. Even with limited training resources, TurnGate achieves a non-trivial improvement over existing defenses. At budget 5 in Figure 3, it reduces ASR from approximately 0.77 with Intention Analysis to below 0.68. Figure 4 shows that these gains persist under distribution shifts.

Each search iteration allows five adaptive probing attempts. A budget of five gives the attacker up to 25 attempts to explore alternative paths based on delivered responses. We believe that more boundary-supervised training data could improve robustness under sustained probing. With additional resources, we can generate more attack trajectories and annotate their intervention boundaries. This would broaden the range of harmful objectives and attack strategies covered during training. We can also generate attacks against the evolving defender and use the newly discovered failure cases for further training.

Q1. Information before the boundary and attacker prior knowledge

An attacker may already possess relevant knowledge and use an earlier response to complete a harmful objective. The defender cannot observe or control that private knowledge. Consider two users asking the same question about a chemical reaction. One is conducting legitimate research, while the other has additional knowledge and needs this answer to complete a harmful process. If their observable dialogues are identical, the defender cannot distinguish between them. Blocking the latter request would also block the legitimate one.

If the defender were required to block any response that could complete a harmful objective when combined with arbitrary private knowledge, many legitimate requests would have to be refused. The AI system can control the information it releases. We therefore aim to prevent the information accumulated in the dialogue from becoming sufficient to realize a harmful objective. Under this criterion, blocking at t∗t^* prevents the first sufficient prefix from being delivered while preserving the preceding dialogue.

PASS before t∗t^* does not imply that the information is risk-free. As discussed in Appendix F.1, an attacker's private knowledge can move the effective boundary earlier.

Q2. Benign pass-through in online OOD evaluation

Thank you for suggesting the online benign evaluation. We added dialogue-level benign pass-through rates for GPT-5.2 and the held-out Gemini target below.

We used 20 fixed tasks from the safe split of XSTest [2] per target. All methods start from the same tasks, with follow-ups conditioned on delivered responses.

TargetReweighted-SFT: dialogue all-pass ↑\uparrowTurnGate: dialogue all-pass ↑\uparrow
GPT-5.218/20 (90%)18/20 (90%)
Gemini-3.1-Pro-preview17/20 (85%)17/20 (85%)

The remaining benign blocks are over-refusals, a central failure mode our evaluation measures alongside harmful interception. This problem is more pronounced for the trajectory-level Naive-SFT baseline in Table 1, which preserves only 60.9% of benign dialogues, compared with 83.4% for TurnGate. The online rates above measure dialogue pass-through. References

[1] Wei et al. The Trojan Knowledge: Bypassing Commercial LLM Guardrails via Harmless Prompt Weaving and Adaptive Tree Search. arXiv preprint, 2025.

[2] Röttger et al. XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models. NAACL, 2024.

We thank the reviewer for recognizing the importance of timely intervention and the value of our benign hard negatives and adaptive evaluations. We appreciate the valuable feedback and would be happy to provide further clarification if any concerns remain.

Thank you again for your constructive review. We have conducted additional experiments, including the requested online Reweighted-SFT comparison and benign evaluation. We also clarified the motivation and limitations of our intervention boundary and discussed robustness under larger attack budgets. We hope these results help address your concerns and look forward to further discussion. If you find our responses satisfactory, we would greatly appreciate your reconsideration of the assessment. Thank you 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.