acceptodds
Under review as a conference paper at ICLR 2027

Multi-Turn Jailbreaks Steer Along Directions Single-Turn Safety Did Not Shape

Abstract

Multi-turn jailbreaks can spread harmful intent across messages that are not individually refusable, yet most deployed defenses score one turn at a time. Across five models, the multi-turn drift direction lies almost entirely outside the plane spanned by single-turn refusal and harmfulness directions. Held-out safety directions provide a positive control, retaining 0.79–0.96 of their norm in this plane; the drift direction retains only 0.03–0.09, consistent with its shuffled-label null on all five models. Directional Context Drift (DCD) reads each user turn with and without history and projects the normalized difference onto a learned attack-mode direction. This difference is zero at turn one, before history exists. A multi-scale EWMA chart with a conformal limit converts these scores into alarms. DCD adds one forward pass per turn and uses the nominal false-alarm rate as its only deployment setting. With an axis for each known mode, DCD reaches 0.957–0.971 AUROC and 0.715–0.879 TPR at 5% FPR averaged over five multi-turn attack families, versus 0.766 and 0.352 for the strongest of three guards. When attacks are paired with benign conversations receiving nearly identical guard scores, the guard sits near chance by construction while DCD retains 0.90–1.00 AUROC. On judge-confirmed attacks whose harm begins after turn one, DCD flags 76–86% before the first harmful response, versus 30% for the strongest guard. Coverage is mode-specific, so attacks steering beyond the learned directions need a new axis.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multi-turn jailbreaks, concept drift, LLM safety, refusal direction, jailbreak detection

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_ZeBBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies internal representation dynamics in multi-turn jailbreaks and proposes Directional Context Drift (DCD) for early detection. DCD compares the hidden state of the current input with and without conversation history, learns attack-related drift directions from labeled malicious and benign conversations, and combines a Direction Bank, multi-scale EWMA, and conformal calibration for online detection. The paper reports strong detection and early-warning performance on known attacks, while also showing limitations under unseen attack patterns and shifts in benign conversational distributions.

Critical strengths
  1. The paper studies the effect of conversation history on internal model representations, rather than relying only on text-level safety classification or refusal behavior. Its observation that multi-turn attack drift directions have low geometric overlap with established refusal and harmfulness directions offers a useful perspective on why single-turn safety representations may not capture multi-turn attack dynamics.

  2. DCD estimates attack-relevant directions through a simple difference-in-means procedure without additional gradient-based training, and combines these directions with established sequential monitoring techniques. The framework is relatively interpretable, and the evaluation includes multiple attack datasets and analyses of context substitution, directional ablations, open-set detection, and adaptive attacks.

Critical weaknesses
  1. DCD learns its drift directions from labeled attack samples, whereas many general-purpose guardrail baselines operate in a zero-shot manner. Thus, the performance gain may partly reflect supervised adaptation to known attack distributions rather than the independent benefit of contextual differencing. This concern is reinforced by Appendix Table 10, where Isolated/Contextual Readout variants can achieve performance comparable to or slightly better than Context Differential. The paper therefore does not fully isolate the value of the additional context-free forward pass and the differential representation under matched supervision and end-to-end detection settings.

  2. The Direction Bank is inherently limited by the attack directions represented in its training data. Novel attacks may induce directions outside the bank, and adaptive white-box attackers may seek to evade known directions. Meanwhile, structure-matched benign dialogues and sensitive-topic conversations can also produce strong drift signals and increase false positives. Recalibrating thresholds may reduce false alarms but can lower attack recall. Consequently, maintaining both high recall and a low false-alarm rate in open-world settings remains insufficiently established.

Questions for authors

Q1. The paper constructs a Safety Plane from the Refusal Direction and Harmfulness Direction and uses low projection onto this plane to motivate its claim about multi-turn attacks. How representative is this selected subspace of the model’s full safety mechanism? Could the authors clarify whether the results support a broad mechanistic claim or only show low overlap with the particular directions selected? Q2. Given that Appendix Table 10 reports comparable performance for Isolated/Contextual Readout variants, could the authors compare Isolated Readout, Contextual Readout, and Context Differential under identical data splits, supervision budgets, layer selection, EWMA monitoring, and threshold calibration? Q3. How does DCD generalize to unseen attacks and shifts in benign conversational distributions? Could the authors evaluate direction-discovery or update mechanisms that do not require labeled examples of each new attack, and quantify how expanding the Direction Bank affects cumulative false positives, detection recall, and calibration?

4 comments

W1/Q2. Table 2 already includes a supervised LAD-based baseline. DCD leads on both AUROC and TPR in 21 of 25 target–family comparisons, and on AUROC in all ten twin comparisons (Appendix J). Both methods use labels, so the difference comes from what they read.

The Isolated/Contextual readouts are single-read variants of our monitor, sharing its estimator, chart and calibration. The new Table 3 compares all three readouts using the same data splits, supervision budgets, procedure for selecting layers, EWMA monitoring and threshold calibration (Appendix E).

With thresholds calibrated on WildChat conversations, DCD flags 24% of benign twins against 38–40%, and 74% of attacks against 77–84% (Table 3). These results show that the second read isolates context-induced drift and enables DCD to considerably lower the false positive rate. The differential is zero before history exists and the raw projections of the readouts already separate attack openers at turn one (0.72–0.78 AUROC, Table 12, Table 10 in the submission).

Q1. The plane combines two established safety-related directions for recognizing harmfulness and producing refusal, supported by causal interventions in (Zhao et al., 2025) and (Arditi et al., 2024). This provides a principled reference for asking whether representations associated with single-turn safety also capture multi-turn context drift.

Refusal can span multiple directions, which motivates extending the analysis beyond the two-direction plane (Wollschläger et al., 2025; Pan et al., 2025). We validate this reference using safety directions estimated from disjoint prompts and extend it to 48-dimensional subspaces built from harmful–harmless representation shifts. These subspaces capture the held-out safety directions well, while most attack-associated drift remains outside them. Subspaces derived independently from base-to-instruct changes support the same finding (Appendix G).

Our conclusion then extends beyond the particular two directions selected. It identifies a geometric gap between measured single-turn safety representations and multi-turn drift, without assuming that these subspaces describe the model’s entire safety mechanism.

W2/Q3. We explicitly evaluate coverage limits, as emphasized by (Sommer and Paxson). With the tested family's axis removed, the remaining directions outperform all three guards in 22 of 25 comparisons, averaging 0.862 AUROC against 0.763 for the strongest guard on the same conversations and five-turn budget. This projection-based test still uses examples of that family for layer selection.

We evaluate adaptive attackers with partial and full knowledge of the monitor (§5.6, Appendix K). At 32 rewrites per turn, DCD detects 76–87% of attacks on Llama under partial knowledge. On Mistral, none of these tested attackers improves joint success over the original attacks (Table 6). Full white-box access enables stronger evasion which we quantify across query budgets in Appendix K, but DCD still performs attack detection for at least half of the attacks.

Appendix H (Table 17) tests growth from one to five axes, recalibrating the shared threshold after each addition. Mean false alarms stay at 5.1–5.8%, while recall across all five families rises from 40% to 86%. We also report recall losses on previously covered families.

Tables 3–4 report benign-twin and sensitive-topic WildChat false alarms alongside attack recall. Table 4 includes all three guards and shows the recall tradeoff before and after recalibration.

The present method learns directions from labeled examples. Reliable automatic discovery and validation of new directions from unlabeled traffic remain future work.

Dear Reviewer,

We wanted to kindly follow up on our rebuttal. With the discussion period ending tomorrow, we would appreciate any feedback and would be happy to clarify any remaining concerns.

Thank you for your time.

Reviewer_ZeBBOpenReview ↗

Official Comment by Reviewer ZeBB

Dear Authors,

Thank you for the clarifications and additional experiments. The matched comparisons and safety-subspace clarification address several of my concerns. I still have some questions about unseen-attack generalization, since examples from the held-out family are used for layer selection.

Given these remaining questions, I believe my original score remains appropriate, and I will retain it.

Thank you for the reply. We are glad that the matched comparisons and safety-subspace clarification address several of your concerns.

The missing-axis test evaluates coverage by existing directions. The layer is selected on separate conversations from the evaluated family, using only the remaining axes. No direction fitted on that family is used to score the held-out conversations in the missing-axis condition (Appendix H). Comparing the same conversations with and without the axis isolates what the family’s direction adds at the selected layer.

The bank-expansion experiment then tests how adding directions changes coverage and false alarms. Each axis keeps its previously selected layer, and the shared threshold is recalibrated on benign chat. Mean recall across all five families rises from 0.403 to 0.860 as their axes are added, while mean false alarms remain at 5.1–5.8% (Table 17).

These experiments characterize the coverage gap and evaluate how to address it. Alignment with the remaining bank is associated with detection without the family’s axis (r = −0.763, Table 15), and adding that axis raises target-averaged AUROC to 0.980–1.000 in the projection-based comparison. Separately, Appendix F shows that 40 attack conversations per axis bring detection before harm within seven percentage points of the full-corpus result for Crescendo and Echo Chamber across three targets.

The abstract explicitly identifies mode-specific coverage as a limit of DCD. Our experiments quantify that boundary and provide evidence about maintaining coverage once examples of a new mode become available. The matched detection comparisons and safety-subspace findings are established independently of the missing-axis experiment.

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

This paper studies how multi-turn jailbreaks show up in model activations and proposes Directional Context Drift (DCD), which reads each user turn with and without the conversation history and projects the normalized difference onto an attack-mode direction learned by difference of means. A multi-scale EWMA chart with a conformal threshold then turns these scores into alarms. In experiments on five attack families and five target models, DCD reaches 0.957–0.971 AUROC and flags 76–86% of successful attacks before the first harmful response.

Critical strengths
  1. The paper is well motivated and the topic is interesting: most existing guards score one turn at a time, while this paper studies how the attack shows up in the model's representations across turns.
  2. The results are strong, especially when no single turn is overtly harmful: on SafeMTData, rescue reaches 0.78–0.92, compared with 0.08 for the strongest guard.
  3. The detector is simple and cheap, needing only one extra forward pass per turn and no gradient training.
Critical weaknesses
  1. The title seems to overclaim. It says multi-turn jailbreaks "steer along" these directions, but Section 4.4 shows that editing the drift coordinate does not change harmful compliance.
  2. There is no human evaluation or second-judge validation. The success labels and the harmful onset all come from one judge and its reliability is not evaluated.
  3. DCD needs white-box activation access and a labeled set for every attack mode, so it cannot be applied to closed-source models such as ChatGPT or Claude without provider access, and it is limited against new attack styles (e.g., cipher drops to 0.62 AUROC on average without its own axis, Table 13).
  4. The adaptive attack is not strong enough to support the robustness claim. It samples at most 32 rewrites per turn, and on Llama detection is still dropping as the budget grows (from 0.873 to 0.711, Table 20).
  5. Figure 1 compares attacks with benign conversations from a different generator; against same-generator twins, the separation drops from about 1.0 to 0.79–0.86 AUROC (Appendix E). The projection analysis in Table 1 is not repeated with same-generator twins, so it is unclear whether the off-plane geometry is partly a generator artifact.
  6. The gain over guards mostly comes from supervision: a fitted probe on the plain contextual read already matches DCD (0.868 vs. 0.862, Table 10), but Tables 2–3 only compare against zero-shot guards.
  7. The abstract and introduction are very number-dense, and many key details are spread across a 28-page appendix, which makes the logic of the paper hard to follow.
  8. One minor thing is the paper uses a lots phrasing such as "rather than ...", which hurts a bit readability.
Questions for authors
  1. Which benign set is used to fit the drift direction in Table 1: the different-generator conversations used in Figure 1, or the same-generator twins? If it is the former, how do the authors rule out that the off-plane geometry reflects generator style rather than attack intent?
  2. Given that editing the drift coordinate does not change harmful compliance in section 4.4, why does the title say that multi-turn jailbreaks "steer along" these directions?
  3. Table 10 shows that a fitted probe on the contextual read reaches almost the same AUROC as the differential (0.868 vs. 0.862). What is the practical advantage of the differential for detection, and why are such fitted probes not included as baselines in Tables 2–3?
  4. How was Granite Guardian validated as the judge for success labels and harmful onset, and why did the authors choose a judge that is also one of the compared guards?
  5. On Llama, detection under the white-box attacker is still decreasing at 32 rewrites per turn (Table 20). What happens with larger budgets, and at what point do the authors expect detection to level off?
5 comments

W1/Q2. “Steer along” describes how attack history changes representations. A direction can track this movement without itself controlling harmful compliance. Section 4.4 clarifies this distinction. We recognize that the title can invite a causal interpretation and propose “Off the Safety Plane: Context Drift in Multi-Turn Jailbreaks” for the camera-ready version.

W2/Q4. We use Granite Guardian’s response-harm assessment, evaluated on BeaverTails, SafeRLHF and XSTest-RH by (Padhi et al., 2025). Its guard role scores user requests, while its judge role assesses generated replies. Our main detection comparisons use corpus labels, so Granite does not provide the ground truth for its own guard comparison. Our geometric findings also do not depend on its response-harm judgments. As a second judge, we used Llama-3.3-70B to assess 743 replies in 100 held-out conversations from SafeMTData, SafeDialBench, cipher, Crescendo and Echo Chamber across all five targets. For each model–family pair, we randomly sampled two successful and two unsuccessful attack conversations according to Granite. The second judge used the same harm definition and current user–assistant exchange, without seeing Granite’s labels. Sampling-weighted reply-level agreement was 93.1%. This complements Granite’s published benchmark evaluation with an independent check of response-harm judgments on our multi-turn conversations.

W3. DCD can use an open surrogate without target activations. On four open-weight targets treated as black boxes, fitting on target conversation text gives mean recall within 1.4 percentage points of direct monitoring. Appendix M (Table 25) also evaluates deployment without target-specific fitting. Generalization to unseen attacks is a longstanding challenge in learning-based security detection (Sommer and Paxson). We evaluate missing-axis coverage (Table 15). An additional comparison on the same conversations and five-turn budget shows the remaining directions outperforming all three guards in 22/25 cases, with mean AUROC 0.862 versus 0.763 for the strongest guard. Cipher identifies a coverage gap. This projection-based test still uses examples of the evaluated family for layer selection.

W4/Q5. We extended every adaptive attacker to 64 rewrites per turn, twice the previous maximum (Table 24). The results are positive. Against the full white-box attacker, the detection on Mistral stays essentially flat at 0.605, with joint success of 0.368. On Llama detection moves to 0.521 with joint success of 0.451.

Broadly, robustness against adaptive attacks remains an open problem in the literature. (Nasr et al., 2026) report adaptive attacks success rates above 90% against most of 12 evaluated defenses, including input detectors. (Andriushchenko et al., 2025) show that adaptive attacks can overcome leading safety-aligned models, with effective strategies depending on the target model. This model dependence is consistent with the different behavior we observe on Llama and Mistral.

For multi-turn detection specifically, the concurrent multi-turn detectors and guards discussed in our paper have not been evaluated against attackers explicitly adapted to their monitors. DCD provides the first such evaluation and shows evidence of robustness under the tested conditions: even with full monitor access and 64 rewrites per turn, it detects more than half of the evaluated attacks on both targets.

These white-box results assume an attacker with an exact copy of the monitor. Resistance is stronger against the tested, more realistic attacker with incomplete knowledge: at 64 rewrites, their joint success is 0.155–0.282 on Llama and 0.158–0.276 on Mistral, compared with 0.451 and 0.368 under full knowledge. With no monitor knowledge, rewriting increases joint success by about 6% on Llama and reduces it on Mistral, relative to original attacks.

W5/Q1. Table 1 used a shared benign reference matched to the attacks’ escalating five-turn structure. The additional analysis with same-generator twins retains only 0.015–0.111 of the drift direction’s norm in the safety plane across all five models (Table 1, Appendix G), supporting the same off-plane finding.

W6/Q3. Table 2 already includes a supervised LAD-based baseline. DCD leads on both AUROC and TPR in 21 of 25 target–family comparisons and on AUROC in all ten twin comparisons (Appendix J). The fitted probes are single-read variants of our monitor, sharing its estimator, chart and calibration. New Table 3 compares them under matched supervision. In Table 12, their raw projections already separate attack openers at turn one (0.72–0.78 AUROC). The second read isolates context-induced drift; the differential is zero before history exists. At ordinary-chat-calibrated thresholds, DCD flags 24% of benign twins against 38–40%, and 74% of attacks against 77–84% (Table 3).

W7/W8. We have revised the paper for clarity, reduced numerical detail, and removed repetitive phrasing.

Thanks for the detailed response.

The same-generator results in Table 1 and Appendix G address my concern about the generator artifact.

My main concern is still whether the differential itself helps detection. In the new Table 3, with the same supervision, the isolated variant has higher AUROC and higher recall. DCD has a lower twin FPR, but it also catches fewer attacks, so this may just be a different operating point.

The B=64 results make me a bit more worried about the adaptive attack. On Llama, white-box detection keeps dropping (0.711 at B=32, 0.521 at B=64), but Section 5.6 says the appendix "confirms the observations, even for higher number of rewrites". I think the main text should show the B=64 numbers and say this directly.

Therefore, I'll keep my current score (Rating: 2: Weak rejection) for now.

Thank you for the follow-up. We are glad the same-generator results resolve the generator concern.

Value of the differential.

The isolated and contextual readouts are variants of our monitor and share its estimator, training setup and alarm rule. All three thresholds are calibrated on ordinary chat at the same nominal rate. At these operating points, DCD flags 24.2% of benign twins against 38.0% for the isolated variant, with attack recall of 74.0% against 77.3% (Table 3). The twin false-alarm rate is therefore 13.8 percentage points lower (36% in relative terms), while recall is 3.3 percentage points lower (4%). We agree that this is an operating-point tradeoff.

What the second read adds beyond this is a measurement. Multi-turn jailbreaks exploit conversation history, and the differential measures how that history changes the representation of the same request. Each isolated score reads the current message without its preceding conversation. Although the monitor accumulates these scores across turns, it does not explicitly measure how history changes the current request’s representation. The twins use benign goals and therefore do not hold that request fixed. Before any history exists, the single-read directions already reach 0.777 and 0.724 AUROC, while the differential is 0.500 by construction. In the raw-projection comparison over turns 2–6, the differential retains similar AUROC (0.862 against 0.877 and 0.868, Table 12).

We test this measurement by holding the final request fixed and replacing its history with a real benign conversation. Harmful compliance falls by 25–44 percentage points, and firing of the single-axis DCD monitor falls from 88–100% to 6–15% (§4.3, Table 20). The isolated representation, and hence its raw score for the final request, remains identical although the model’s behavior differs. The contextual read also responds to the replacement. The differential’s role is to measure the change relative to the same request presented without history.

Similar AUROC does not resolve our question about the representation change contributed by history. The differential isolates that change, allowing us to attribute the observed geometry to history. The change remains largely outside the tested single-turn safety subspaces, even when expanded to 48 dimensions, and the same-generator controls support this finding (Table 1, Appendix G). This geometric analysis, together with the detection framework and its measured tradeoffs, is the paper’s contribution.

Adaptive attacks.

We now report the B = 64 results in the main table and explicitly state in Section 5.6 that white-box detection on Llama decreases from 0.711 at B = 32 to 0.521 at B = 64, while joint success increases from 0.250 to 0.451. We also distinguish this trend from Mistral, where the corresponding rates remain approximately stable.

We have removed the statement that higher budgets “confirm the observations.” The revised discussion acknowledges that, although DCD still detects more than half of the rewritten conversations on both targets at the largest tested budget, detection can decrease substantially as the budget grows depending on the target. We have also updated the Limitations section to acknowledge this model and budget dependence. The complete results across budgets remain in Appendix K.

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

Most safety guardrails look at one message at a time, meaning that multiple individually innocuous messages can be put together to produce harmful behavior that the guardrail misses. This paper addresses this problem of multi-turn jailbreaks with two main contributions.

First, the authors look at the drift of multi-turn jailbreaks: the difference in direction in activation space between user messages with the full context of the conversation (i.e. the full trajectory that would be harmful) vs without that full context (just the user message). Then, they compare this drift direction with the single-turn harm-refusal plane and find it lies outside the span of that plane.

Second, the authors propose Directional Context Drift (DCD), a new detection method for multi-turn jailbreaks that accounts for this drift. They validate this with experiments across five models.

Critical strengths
  1. The setup of diagnosing the problem (drift lying outside span of the harm-refusal plane) motivates the DCD method well, making it intuitive what the method is doing.
  2. The experimental setup covers a reasonable number of models, and the method detects harmful multi-turn behavior across these models.
Critical weaknesses
  1. My primary concern is with the significance of the contribution. First, it would be better to test this in more deployed settings where false alarms are meaningful. As I understand, it is good at detecting harmful behavior in multi-turn jailbreak settings, at the cost of some increase in false alarms ("Third, sensitive-topic WildChat raises false alarms, to 0.17–0.20 on Llama-3.1-8B against 0.035–0.050 on held-out chat.") I am aware Appendix J also has some results on false alarms. These would be good to put in the main text, since it's an important tradeoff the method has to make. Second, the authors mention that the drift axis is not "a causal jailbreak direction" (line 310). As I understand, this means it does not isolate the actual cause of this behavior in the model, so we can't meaningfully intervene on this direction to actually reduce harmful behavior, only detect it.

  2. The story of the paper is hard to follow, as the paper lacks meaningful transitions between subsections and sections. It would greatly improve the readability of this work if i) the authors added more transitions ii) explained the higher-level questions and takeaways of different subsections instead of relying primarily on numerical results in large parts of the paper.

Questions for authors
  1. Can the authors clarify more on the the tradeoff of the method with respect to false positives? How meaningful is this?
  2. This method clearly detects harmful behavior better than the baselines, but at the same time this direction is not causal. Can the authors speak on what this tells us about the cause of harmful behavior in the model, or even about detecting it more broadly?
3 comments

W1.

Significance of the contribution: Our main contribution is an analysis and geometric diagnosis of multi-turn jailbreaks through context drift. Despite its importance for safety, the relationship between history-induced drift and single-turn safety representations remains underexplored. By isolating the representation change induced by conversation history, we show that attack-associated drift remains largely outside measured single-turn safety subspaces, even when expanded to 48 dimensions (Appendix G). DCD demonstrates how this history-dependent signal can support detection before harmful responses.

False alarms on realistic benign traffic: We have moved the false-alarm comparison from Appendix J into §5.3 (Table 4), including all three guard baselines. The evaluation uses real sensitive-topic WildChat conversations and reports false alarms and attack recall before and after recalibration. DCD’s mean false-alarm rates after recalibration are 5.2–6.1%, while its recall exceeds the strongest guard in nine of ten target–family comparisons. The table makes the recall tradeoff explicit; detailed results remain in Appendix J.

Detection and causality: DCD measures history-dependent changes and provides early warning. Replacing the attack history while keeping the final request fixed reduces harmful compliance by 25–44 percentage points and lowers DCD’s alarm rate (Table 20). Our interventions do not establish that the drift coordinate itself controls harmful compliance. Its practical value is early detection. DCD flags 76–86% of successful multi-turn attacks before their first harmful response, when an alarm could trigger blocking or further review. We clarify this distinction in §4.4.

W2. We have revised the paper for clarity, reduced numerical details, added more transitions and added more questions and takeaways to introduce and close sections.

Q1. The tradeoff differs across target models. On Mistral, recalibration leaves recall nearly unchanged (82.8% to 81.9%). On Llama, it reduces false alarms from 18.4% to 5.2%, with recall falling from 82.3% to 66.7%. The benefit is fewer legitimate conversations flagged, with a recall cost that must be assessed for each target. Table 4 reports both operating points. After recalibration, DCD still achieves higher mean recall than all three guards.

Q2. Our analysis locates attack-associated history effects largely outside the measured single-turn safety subspaces. The history-replacement experiment shows that changing the preceding conversation changes behavior even when the final request stays fixed (Table 20). The coordinate-editing experiments do not establish which internal mechanism causes that behavioral change. For detection, the broader implication is that history-dependent representation changes provide useful evidence of risk, even when editing the measured direction has not been shown to prevent harmful compliance.

I thank the authors for their effort in rewriting the paper. The readability has improved, and I have adjusted my “Presentation” score. I’m still concerned about significance of the contribution. If analysis + geometric diagnosis is the primary contribution, I would expect more insight into how these jailbreaks work beyond detection. The analysis shows that the drift lies outside the measured plane, but most directions in high-dim spaces will lie outside this low-dim subspace. This lines up with Table 1 showing the projection is similar to the shuffled-label null. It is confusing why this is considered a finding or significant, when it is the default (that any direction is outside the span and therefore similar to the null). Therefore, it is confusing what is significant about the finding. I have updated to a very weak 3.

Thank you for raising both the presentation score and the overall rating. Your comments on readability improved the paper significantly.

The shuffled-label null in Table 1 measures the high-dimensional effect you describe. It shows how far an unrelated direction lies from the plane. The same projection measure gives a high value for directions related to safety. Held-out refusal and harmfulness directions retain 0.79–0.96 of their norm in the plane. The attack direction retains only 0.03–0.09 of its norm in the plane, showing no detectable alignment with the refusal–harmfulness plane beyond that observed for directions fitted with shuffled labels. The result holds with same-generator benign twins. We also expand the plane to 48 dimensions that capture most of the harmful–harmless shift variance beyond it, yet most of the attack direction remains outside (Appendix G). These safety subspaces therefore capture established safety features but largely miss the representation change induced by attack history. This contrast is the finding.

Refusal suppression has been demonstrated in single-turn jailbreaks (Arditi et al., 2024). If the multi-turn attacks we study worked the same way, the refusal readout would fall as the attack proceeds. It stays within 0.25 of its benign reference while the drift readout rises to 3.3 benign standard deviations (§4.3).

What this tells us about the attacks is that they progress while the measured refusal readout remains near its benign reference. This is why we monitor the change induced by history.

We hope this clarifies why we consider the result significant.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and additional experiments

We thank the reviewers for their careful reading and constructive feedback and the Area Chair for coordinating the review. Their comments have helped us clarify the contribution and strengthen its evaluation.

Our main contribution is a geometric analysis of how conversation history changes model representations during multi-turn jailbreaks, with DCD demonstrating the detection value of this signal. We have made the following revisions and additions.

  • Clarified the distinction between history-induced representation changes and causal control of harmful compliance.
  • Brought the matched-supervision and sensitive-topic false-alarm comparisons into the main text, reporting attack recall alongside false alarms.
  • Added same-generator controls for the geometric finding, surrogate-monitoring results, and an analysis of bank expansion and recalibration.
  • Conducted a second-judge consistency check and extended the adaptive-attack evaluation.

We have also improved the presentation and highlighted new and revised text in blue to make the changes easy to identify.

  • Simplifying the abstract and introduction, reducing numerical detail and clarifying the contributions.
  • Revising section headings, transitions and takeaways to make the argument easier to follow.
  • Adding a method overview diagram showing how the two reads lead to an alarm.
  • Adding a navigation table connecting the main analyses to their supporting appendices.
  • Simplifying wording and reducing repetitive phrasing, including “rather than.”

Our separate point-by-point responses address each reviewer’s concerns and explain the new results and their scope, with references to the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.