acceptodds
Withdrawn from ICLR 2027

Jailbreak Context Lingers: Divergent Safety Routing and Its Cross-Task Predictability in Tool Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

As large language models increasingly operate as tool-using agents, post-jailbreak safety feedback is often assumed to serve as a reliable safeguard; however, how lingering jailbreak context shapes subsequent agent behavior remains largely unexplored. To systematically examine this dynamic, we introduce a paired continuation framework across 192 parent tasks spanning 42 domains, evaluating 12,148 analyzed continuation pairs (curated from a 12,288-pair initial design) across eight diverse agents. We find that identical safety feedback induces sharply model-dependent behavioral routing rather than uniform protection: redirecting unsafe trajectories toward legitimate completion (*rescue*), sustaining unauthorized execution (*persistent unsafe*), or triggering over-refusal on benign tasks (*collateral loss*). Through layer-wise activation patching, we discover a shared *late-commit pattern* where causal intervention effects surge sharply near the final layers (relative depths of 0.958–0.984) despite an over 30-fold variation in peak magnitude across architectures. Crucially, critical-layer representations correlate with macroscopic routing outcomes, and intervening at these layers causally alters concrete next-step tool actions. Building on this causal foundation, we test whether localized intervention-derived features can serve as predictive proxies for full-trajectory routing outcomes on unseen parent tasks under leave-one-parent-task-out evaluation, finding that they provide viable predictive signals in responsive agents with peak ROC AUCs reaching 0.675 for *rescue*, 0.777 for *collateral loss*, and 0.702 for *persistent unsafe*. These findings establish a mechanistic lens and a predictive baseline for anticipating the safety and utility trade-offs of post-jailbreak feedback in autonomous agents.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Tool-Using Agents, Jailbreak Persistence, Mechanism

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_G4M3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Safety feedback delivered mid-trajectory is often assumed to protect a jailbroken tool-using agent. This paper tests that assumption with a paired continuation design: agents are resumed independently from identical frozen pre-action checkpoints, once under a neutral reminder and once under safety feedback, so the multi-turn history is held exactly fixed. Across many parent tasks and eight open-weight agents, the authors find that identical feedback produces sharply model-dependent outcomes — rescuing some unsafe trajectories, leaving others unchanged, and inducing over-refusal on benign tasks. They then apply layer-wise activation patching, report that intervention effects concentrate near the final layers, show that patching at those layers changes the next tool call, and test whether features from those layers predict trajectory outcomes on held-out tasks.

Critical strengths

The paired continuation design is careful experimental work and it is what gives the behavioural results their force. Resuming independently from the same frozen checkpoint under two conditions, with deterministic replay, verified simulator copies, greedy decoding, and fixed budgets, isolates the effect of the feedback from everything else in a long multi-turn trajectory. That is harder to do well than it sounds, and the validation step — confirming on independent simulator instances that the legitimate workflow remains completable without the unauthorised action — rules out a confound that would otherwise make rescue and collateral loss impossible to distinguish from task infeasibility.

The three-way outcome taxonomy is the second strength, and I think it is the paper's most durable contribution. Including collateral loss on benign variants, alongside rescue and persistent unsafe, captures a trade-off that safety-only evaluation designs structurally cannot see: the same intervention that saves one trajectory breaks another. The resulting behavioural finding stands independently of everything mechanistic in the paper. Persistent-unsafe rates vary by nearly two orders of magnitude across agents and do not track parameter count, and rescue and collateral loss peak in different model families entirely. For anyone designing runtime safety feedback for agents, that result alone is worth knowing, and it is established by the behavioural half of the paper without relying on the patching analysis at all.

Critical weaknesses

The late-commit pattern is the paper's headline mechanistic claim, and the metric used to establish it is confounded in a way the paper anticipates but does not neutralise. Replacing a residual-stream state at layer ℓ and measuring cosine displacement of the resulting next-token logits will mechanically produce larger effects as ℓ approaches the output head, simply because fewer remaining layers can absorb the substitution. Seven of eight curves peaking at the final layer is exactly the profile a pure structural artifact would produce. The paper offers two rebuttals — vocabulary-space projections separating routing outcomes, and argument-level tool reversals that mid-depth controls fail to achieve — but both of these also benefit from proximity to the output, so neither isolates the effect from the confound. The fix is a control contrast rather than a new method. Run the identical patching sweep on a pair that differs only in surface wording and carries no safety content, such as two neutral reminders phrased differently. If that control produces the same late ramp, the pattern is structural and the claim should be reframed; if it stays flat while the safety contrast surges, the functional interpretation is earned and the paper becomes considerably stronger. Normalising the safety curve against that control, or patching at a fixed offset from the final layer to make depths comparable across architectures, would serve the same purpose.

The second concern is that the predictive results are reported by their maxima when the distribution does not support that summary. The three peak values quoted in the abstract each come from a different agent, and the per-agent table contains several cells at or below chance. More importantly, some estimates rest on very few positives — one rescue AUC is computed from four positive examples, and one persistent-unsafe AUC from three — and the paper notes this as high sample variance but then reports the peaks as the finding. This is readily fixable and I think the honest version is still publishable. Report confidence intervals in the main table, suppress or visually mark cells below a minimum positive count, and summarise with a median or aggregate across agents rather than a maximum. If the resulting statement is that intervention features carry modest predictive signal in some agents and none in others, that is a useful and credible baseline, and it is consistent with the paper's own framing of post-jailbreak feedback as model-dependent. As a smaller matter, the abstract and the first contribution bullet give different pair counts, and the phrase describing the curated set is ungrammatical; both are easy to correct.

Questions for authors

The question that would most change my assessment concerns the patching control. Do you have, or can you run, the layer-wise sweep on a non-safety contrast — two neutral reminders differing only in wording — on even two or three of the eight agents? If the late-layer surge is specific to the safety contrast, that result belongs in the main text, and it would substantially strengthen the mechanistic half of the paper.

I have two clarifications. First, could you report confidence intervals alongside the per-agent prediction AUCs, and state a minimum positive count below which an estimate should not be interpreted? I would like to know how many of the reported cells survive such a threshold. Second, the abstract reports one count of analysed pairs and the contributions section another; which is correct, and what was the curation criterion that removed the difference?

1 comment
AuthorsOpenReview ↗

Gentle Reminder: Follow-up on Our Rebuttal

Dear Reviewer, as the discussion deadline approaches, we kindly invite you to review our previous response to your insightful comments. We hope our addtional experimental results and clarifications have fully addressed your concerns. Please let us know if any points remain unclear or require further discussion—we would be glad to provide additional details before the period closes. Thank you again for your time and constructive feedback!!

Reviewer_2urbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how tool-using agents respond to safety feedback after jailbreak attempts. Through paired continuation experiments across eight models, the authors find that identical safety feedback can lead to different outcomes, including rescue, persistent unsafe behavior, and collateral loss. They further use activation patching to identify a late-commit pattern and explore whether intervention effects can predict downstream agent behavior.

Critical strengths
  1. The research problem is interesting and practically relevant. Understanding how agents respond to safety feedback after jailbreak attempts is important, especially when unsafe tool actions can have real consequences.

  2. The paired continuation framework is carefully designed. By comparing different feedback conditions at matched checkpoints, the authors can better isolate the effects of safety feedback on subsequent agent behavior.

  3. The evaluation is fairly comprehensive, covering eight models, 192 tasks across 42 domains, and multiple types of analyses, including behavioral evaluation, activation patching, and cross-task prediction.

Critical weaknesses
  1. The outcome definitions do not fully match their interpretations. The definition of rescue includes cases where the original trajectory simply fails to complete the task, rather than executing an unsafe action. Similarly, collateral loss includes incomplete execution and other failures, not just over-refusal.

  2. The late-commit pattern may be overinterpreted. The authors find that activation patching has the largest effect near the final layers. However, this is somewhat expected, since interventions closer to the output naturally have a more direct influence on logits. Although the tool-action experiments show that these interventions can change agent behavior, they do not establish that safety decisions are actually made in the final layers.

  3. The role of lingering jailbreak context is not entirely clear. Many of the main results combine experiments with retained and neutralized jailbreak contexts. The effect of retaining the jailbreak also varies considerably across models. I think the paper would benefit from a clearer analysis of how much the observed behavioral differences are actually caused by the lingering jailbreak context, rather than general differences in how models respond to safety feedback.

  4. Limited predictive performance and missing baselines. Although the authors report peak AUCs of 0.675, 0.777, and 0.702 for the three outcomes, many results are only slightly above random guessing. Moreover, the paper mainly compares intervention features across layers, without evaluating simpler alternatives such as directly using the models' output logits. It is therefore unclear how much additional predictive value the proposed features provide.

Questions for authors

See weakness

Reviewer_6j7wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how a safety reminder changes tool-agent behavior when jailbreak context is retained or neutralized. Matched continuations across eight agents distinguish rescue, persistent unsafe execution, and collateral loss. Layer-wise activation patching identifies large late-layer effects on next-token logits, and intervention-derived features predict some full-trajectory outcomes under parent-task-held-out evaluation.

Critical strengths

S1: The matched design separates safety from legitimate task completion. Section 3.1 varies feedback and jailbreak context at fixed task checkpoints. The three outcomes avoid treating blocked unsafe actions and successful legitimate work as interchangeable. Section 5.2 holds out every variant of the test parent, and Table 3 reports positive support.

S2: The intervention affects executable tool decisions. Section 5.1 checks tool names and arguments, with half-depth, random-perturbation, and self-patching controls. This connects the representation analysis to behavior rather than relying only on logit geometry.

Critical weaknesses

W1: Safety-specific localization remains unresolved. Section 4 acknowledges output-head proximity, while Section 5.1 demonstrates local causal efficacy for concrete tool actions. However, half-depth and random controls do not distinguish safety-specific localization from generic late-state sufficiency. Equally late, meaningful non-safety instruction contrasts would discriminate these explanations. Section 5.1 also properly limits its endpoint to first-call matching, not terminal safety recovery. The macroscopic-routing interpretation in Sections 5.2 and 6 therefore remains indirect; continuing patched rollouts is a proposed test, not a result the authors claim to have measured.

W2: Incremental predictive value over simpler signals is unclear. The paper presents an initial predictive baseline. Table 3 reports its performance and Figure 7 compares layer choices, but neither tests incremental value against simpler pre-action predictors. Rescue AUC is below 0.60 for six of eight agents; two estimates are below chance. Comparisons against equally accessible unpatched next-action probabilities, tool disagreement, or refusal signals on identical held-out parents, with paired uncertainty, would clarify the intervention feature's incremental value.

Questions for authors
  1. In Sections 4-6, is late-commit intended as a sufficient intervention site for the next tool decision, or as a distinctive location of safety computation governing full-trajectory routing?
  2. Do equally late patches from meaningful non-safety instruction contrasts show the same layer concentration and tool-action effects?
  3. On identical held-out parents, how do intervention features compare with equally accessible pre-action output or tool-disagreement signals, with paired uncertainty?
  4. If the patched first decision is followed through, how often do terminal outcomes change among rescue, persistent unsafe execution, and collateral loss?

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

Related papers

Discussion (0)

Sign in to comment.