acceptodds
Under review as a conference paper at ICLR 2027

Towards A Unified Misuse Monitoring Benchmark

Abstract

LLM agents increasingly act in multi-actor environments, exposing them to misuse from multiple sources: decomposition attacks, where a harmful request is split into innocuous sub-requests, and prompt injection attacks, where a compromised tool delivers a malicious instruction. Existing evaluations treat these threats separately and ask whether a trajectory is harmful, rather than when it becomes harmful. We propose monitoring the agent's responses, where its actions are externalised, and ask whether the first point where monitors identify harm lands within a harm window (from the agent's first harmful commitment to goal execution). We develop a unified formalism for trace-level misuse monitoring and use it to construct a benchmark of  6,200 conversation transcripts between a user, an LLM agent, and the external environment, spanning both threats in a shared schema, with a labelled harm window, corresponding benign controls, and matched instances of refusals to these requests. Across 17 monitor configurations, we find that our proposed action-framed monitors perform well on both threats under classical metrics (AUC: 0.95 and 0.99 respectively), while content-framed monitors collapse on injection attacks (AUC: 0.52). We also show that classical position-blind metrics paint an optimistic picture of monitor performance, since all monitors localise decomposition attacks poorly under the interval metric, which measures the ability to localise harm. Broadly, we illustrate the need for a unified study of misuse monitoring.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
misuse monitoring, LLM agents, prompt injection, decomposition attacks, agent safety, AI control, benchmark, evaluation metrics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZyLsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies misuse monitoring for LLM agents, bringing decomposition attacks and prompt injection attacks into a unified framework for evaluating agent trajectories. A monitor receives interval credit only if its first alert falls within this window. The main finding is that monitors focused on the agent's actions distinguish successful injection from refusal more effectively than monitors focused on input content.

Critical strengths
  1. Distinguishing eventual detection from timely intervention better reflects agent safety requirements than classifying entire trajectories alone. This is an important and practically relevant problem.

  2. The paper evaluates multiple monitoring configurations and models, with additional analyses of calibration, cost, and window sensitivity.

  3. The appendices acknowledge annotation disagreements, generation artifacts, and the influence of evaluation cadence.

Critical weaknesses
  1. The core metric lacks sufficient justification. Reasonable alerts during clearly malicious preparation can receive zero credit. Low interval credit may therefore reflect useful early intervention rather than poor monitoring. Conversely, an alert inside the window may arrive after some harm has already occurred.

  2. The claimed methodological advantage is confounded. Action framing changes both the judgment target and evaluation cadence. Matching cadence raises Gemma's sequential monitor interval credit from 0.166 to 0.410, close to the action variant抯 0.433, weakening attribution of the improvement primarily to framing.

  3. Synthetic traces may make detection unrealistically straightforward. Successful injection continuations are instructed to explicitly announce the malicious action and immediately execute it. Delayed, concealed, and interleaved attacks are insufficiently represented.

Questions for authors

Could you rebuild the refusal twins using the final evidence-onset labels and report whether the conclusions remain unchanged? How sensitive are the timing results to annotation disagreements?

2 comments
AuthorsOpenReview ↗

Response to Reviewer ZyLs (1/2): justifying interval credit and matched cadence

We thank the reviewer for a careful review. We have added some new experiments to Appendix H based on your concerns.

Justifying interval credit

We agree that interval credit is strict. We reported recall next to interval credit throughout the submission, along with the split into early, timely and late flags (§3.4, Figure 4; App. E.1, Figure 10). We also inspected some of the early flags. 18 of 21 fell on preparation, such as drafting a phishing message, so we treat the interval credit as a lower bound on detection (App. E.1). We have now also re-scored every arm with nine different ways to assign monitor credit (App. H.3):

  • Early flags on preparation. An action-framed arm still leads when early flags earn half credit, and keeps the lead until they earn 80%. Many of these are flags well before the harm: a median of 7–9 messages before the request that elicits the first harmful commitment.
  • Flags inside the window but after some harm. App. B already shows how harm can build up inside the window. If we end the window earlier (like at O), it lowers credit by at most 0.07 (App. D.10, Table 12). If we discount flags that come after an earlier harmful tool call, the ranking of the arms barely changes; but the top two arms become statistically tied.
  • For injection attacks, action-framed arms beat content-framed arms under all nine rules.

§3 now presents interval credit as the strictest of this family of rules, and App. H.3 shows recall as the most lenient.

Experiments that match the cadence

As the reviewer notes among the strengths, the submitted appendix already reports a control for cadence (App. D.10, Table 12). We agree with the claim for decomposition, and the paper reaches the same conclusion (App. E.1, §6). There, both framings catch almost the same traces, and the timing advantage of sequential_action comes mostly from scoring only at the agent’s messages.

We have now added this control for injection, where our framing claim matters more (App. H.5). When we match the cadence of arms on injection, the AUC for the (Gemma) sequential arm rises from 0.521 to 0.646, but is still far below sequential_action (0.985). But when we then add one sentence telling the content-framed monitor to judge the assistant's actions (thus changing the judgement target), it recovers most of the remaining gap (86% on Gemma, 91% on Qwen3.6).

So the two changes (changing the cadence vs changing the framing) have separate effects. Cadence mainly changes timing on decomposition. The framing (or judgement target) mainly changes detection on injection. So action-framing allows one monitor to perform well on both settings. §5 now states this more clearly.

AuthorsOpenReview ↗

Response to Reviewer ZyLs (2/2): diversified attacks, rebuilt refusal twins and label sensitivity

More diversified attacks

The submitted version covers part of this. Attacks interleaved with benign steps have their own split (decomposition_hard, §4.1; App. D.6). We had also run an adversarial study that rewrites the finished traces to hide the attacks better (while keeping the harm). This succeeded in at most 5 of 40 cases (App. E.5, Table 17). We now further add two tests:

  1. No announcement. When we delete all the assistant text and keep only the tool calls, then sequential_action achieves 0.969 AUC on synthetic traces and 0.935 on real AgentDojo runs.
  2. Concealed, delayed attacks. We generated traces where the agent never mentions the injection prompt, takes at least two user task steps first, and also keeps its final reply about the user's task. The sequential_action arm's AUC is 0.989 and it flags every attack.

The baseline arm (which sees one message at a time) drops in AUC from 0.938 to 0.846, so it does partly rely on the announcement. The concealment behaviour is also still prompted, so this does not represent an adaptive adversary.

Refusal twins and labels

Our submitted version flagged this issue (App. C.2, "Refusal twins"): the twins were created as per an older version of the onset labels. We have now rebuilt all refusal twins on the final labels, and results have been updated in §6 and App. E. All our conclusions still hold:

  • sequential_action now flags 47–72% of E-twins (refuse immediately at E) on 8 of 9 models (previously 28–57%).
  • The action-framed arms have an AUC of 0.67–0.86 on separating harmful traces from their twins (previously 0.76–0.91).
  • The change comes almost entirely from traces where the onset was updated to be placed later.

We reported rater reliability (App. C.4.1) and a grace-period sweep showing how credit moves if E sits a few messages late in the submission (App. D.10, Figure 9) to address label sensitivity. Now in App. H.6, we find that alternative labellings move pooled interval credit by at most 0.055 per arm (earlier labels), 0.003 (cross-family labels), 0.037 (shifting a random third of labels by one step), and 0.006 (author-labelled subset). Action-framed arms perform better than content-framed ones. The best arm is still the same in every case. Only large systematic shifts of the labels affect the scores substantially.

All revisions are listed in our general response, and revised text is highlighted in blue.

Reviewer_6vEfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The work proposes a formalism that unifies decomposition attacks and prompt injection attacks and evaluates monitoring performance from the perspective of a harm window. The harm window spans from the agent's first commitment to a harmful action to the execution onset of the overall harmful goal. The authors show that although different monitors achieve reasonably high recall in identifying harmful trajectories, they often flag them much earlier than the harm window begins. They additionally find that action-framed monitoring generalizes better across attack types and that monitors frequently flag trajectories that eventually end in refusals.

Critical strengths
  1. A formalism to unify decomposition and prompt injection attacks, with the potential for straightforward integration of additional attack types.

  2. I like the insights extracted from the experiments, particularly: a. Even though monitors have high recall, they often flag harmful trajectories much earlier than the harm window. Moreover, many of these early flags persist even when the subsequent harmful actions are replaced with refusals. b. Framing monitoring around the agent's actions rather than the harmful intent of the inputs improves generalizability across attack types.

Critical weaknesses
  1. Even though this is listed as a limitation, one model, Gemma, generates all the trace continuations and simulates all tool outputs. What measures are in place to ensure that the generated agent responses and follow-up interactions resemble those in a real environment? In particular, the generator is explicitly instructed to produce successful attacks or refusals, rather than these outcomes emerging naturally from the target agent's behavior. Using an automated red-teaming framework like STING (https://arxiv.org/abs/2602.16346) could provide more realistic attack trajectories. A discussion of why the current single-model generation approach is preferable would be useful. Finally, some independent human review of the generated traces would give me more confidence in their realism.

  2. At least a subset of the decomposition evidence-onset labels should be verified by humans. This is particularly important because the proposed metric directly depends on these annotations, while the reported exact-index agreement between different model-family raters is only 67.5% on the 40 double-coded traces. It would also be useful to understand how sensitive the results are to errors in these labels.

  3. The contribution feels somewhat incremental, as the work mainly brings together existing attack datasets and monitor configurations under a unified evaluation formalism. Prior work has also considered the temporal aspect of misuse detection. For example, Brown et al. (2026, https://arxiv.org/abs/2605.31593) explicitly evaluate time-to-detect and demonstrate the benefits of detecting harmful activity earlier. While the precise harm-window metric is different, I would appreciate a clearer discussion of what additional insights it provides beyond existing temporal monitoring evaluations.

  4. I also have some concerns about the conclusions drawn from the refusal study. While it demonstrates that monitors can flag interactions that eventually end in refusal, is this necessarily equivalent to unnecessarily interrupting useful benign work? After all, the underlying requests may still be malicious, even if the agent ultimately refuses to comply, and for refusals occurring after the evidence onset E, the agent may already have performed harmful intermediate actions. Moreover, the refusal traces are constructed by replacing a harmful continuation with a refusal at a predetermined point, ensuring identical prefixes by construction. In real interactions, agents that eventually refuse might behave differently even before that point. How representative are these synthetic refusal twins of naturally occurring refusals, and to what extent can the observed early-flagging rates be generalized to real deployments?

Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6vEf (1/2): realism of generated traces and human label verification

We thank the reviewer for a detailed review. We provide some new experiments in Appendix H to address these concerns.

Realism of the generated traces

A monitoring benchmark needs both harmful and benign outcomes on every task pair, but recorded successful attacks in the real world are too few and too unevenly spread to provide a balanced set. This is why we generate synthetic traces. We stated the single-generator limitation in §6. The submitted version also generates in a constrained format. Every injection trace starts from a recorded run from one of the agents (App. C.3), and the generated completions have to pass integrity checks. We also showed that our generated traces can replicate known phenomena (App. C.5, D.11).

We did not run a human review of realism. Instead, we investigated whether our main conclusions depend on the generator:

  • Outcomes from real agents. On 1,592 real AgentDojo runs from 21 agent models, action-framing still beats content-framing by 0.44–0.48 AUC on both monitors. Real runs are harder, and §6 now reports this (App. H.1).
  • A different generator. When we regenerate the traces with Qwen3.6, each arm’s AUC changes by at most 0.03, and this doesn't affect any conclusions. Neither monitor favours traces written by its own model (App. H.2).
  • Natural refusals from an agent never told to refuse (App. H.8; see below).

We did not run STING. STING is better suited for red-teaming a fixed target. The formalism is meant to take new threat types, and STING sessions could be added as the benchmark covers more threat types, once each session has a labelled point where harm begins (App. A.2). Nevertheless, we now cite STING in App. F.

Human verification of evidence-onset labels

We agree that a subset of traces should be human-verified. One of the authors has now labelled a subset of 40 traces that both LLM raters labelled using the same brief. The author agrees with the blind rater exactly on 25/40 and within one announcement on 28/40. They agree on whether the trace contains harm at all on 37/40. This is close to the agreement between the two LLM raters (27, 30 and 34 of 40; App. C.4.1). App. H.6 gives the details.

To measure sensitivity, we re-scored every arm under five alternative labellings (App. H.6): the earlier labels, a second rater’s labels, the author's labels, random one-step shifts, and a systematic shift of every label. In the first four, interval credit moves by at most 0.06 per arm, and action-framing stays ahead of content-framing. On the author’s labels it moves by at most 0.006. Shifting every label in the same direction moves absolute credit much more, but the best arm is the same in all five.

AuthorsOpenReview ↗

Response to Reviewer 6vEf (2/2): relation to time-to-detect and the refusal study

Contribution and relation to time-to-detect

The revised text now cites Brown et al. in §2. The submitted version stated that no existing benchmark “scores the flag against a window labelled on both sides” (§2.4), and contrasts this with deadline-only credit, which is “position-aware in one direction only” (App. F.3). Time-to-detect rewards any earlier flag. Our harm window keeps a deadline but also adds a lower bound, the agent’s first harmful commitment, which changes conclusions. On decomposition, the arm with the highest recall (sequential, 0.956) has one of the lowest interval credits (0.166) (Table 2). If all early flags earned full credit, then the arms score close to their recall and the timing differences would disappear (App. H.3). The window helps us ask whether a flag is early enough to act on, and the early-flag finding the reviewer highlights comes from that question. For example, the action-framed arms' early flags come at a median of 7–9 messages before the harmful request, on preparation, whereas most early flags of sequential fall on the request itself (§6). The time-to-detect metric would credit them both alike.

The refusal study

We agree that flagging a refused malicious request is not the same as interrupting useful benign work. App. E.2 (“Caveats”) already notes that “the twin still carries the harmful request”. It also counts flags between E and the refusal as detections, not false positives (§6; App. E.2), and its main rate uses twins that refuse at E, before any harmful step. §6 and the ethics statement now say that a flag on such a twin uses audit budget, and whether that is a cost depends on the deployer’s policy. Rebuilt on the final labels, sequential_action still flags 47–72% of these twins on eight of nine models.

On representativeness, we ran our sub-requests through an agent that was never told to comply or refuse (App. H.8). 93% of its refusals came at or before E, so refusals after harmful steps are uncommon. You are right that such agents behave differently before refusing. Action-framed arms flag only 6–19% of natural refusals, close to the 10% budget, so twins that refuse at E are a worst case for them. Content-framed sequential still flags about 90% of natural refusals, so that result does not depend on how the twins are built.

All revisions are listed in our general response, and revised text is highlighted in blue.

Reviewer_EgSBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a unified benchmark for evaluating misuse monitoring in LLM agents, focusing on two attack scenarios: decomposition attacks and prompt injection attacks. The authors argue that existing monitoring evaluations primarily measure whether a harmful trajectory is detected, while overlooking when the detection occurs. To address this limitation, they introduce a harm-window formalism that identifies the interval between an agent's first harmful commitment and the execution of the overall harmful objective.

Critical strengths

Important and interesting research problem. The paper identifies a meaningful limitation of existing agent-monitoring benchmarks by considering when harmful behavior is detected, rather than only whether it is detected.

Comprehensive evaluation and useful observations. The authors construct a unified benchmark covering two different attack scenarios and evaluate a variety of monitoring configurations across multiple LLMs.

Critical weaknesses

The benchmark relies heavily on synthetic, outcome-conditioned trajectories.

The proposed harm-window metric and evaluation protocol may not accurately reflect practical monitoring effectiveness.

The artifact is imcomplete.

Questions for authors

My primary concern is the validity of the synthetic benchmark. The author explains that the agent continuations and tool outputs are generated using Gemma 4. an the model is provided with the overall harmful objective when generating decomposition trajectories, while Appendix C.7.3 explicitly instructs the model to produce successful or unsuccessful prompt injection outcomes. However, in realistic attack scenarios, the agent does not necessarily know the attacker's overall objective, and its behavior is not predetermined. I am concerned that these generation choices may introduce artificial patterns that monitoring models can learn to recognize. Although the authors provide some ablation studies, I do not think these fully resolve the concern. Could the authors evaluate their benchmark findings using actual agent execution trajectories, for example, successful and unsuccessful runs from AgentDojo? This would help establish whether the conclusions generalize beyond simulated interactions.

I am not entirely convinced by the definition of the harm window and its relationship to practical security objectives. The author propose a monitor receives credit only when its first detection falls between the evidence onset and execution onset. However, detecting a suspicious instruction before the agent commits to harmful behavior can also be desirable. For example, if a monitor identifies a malicious prompt injection immediately after the agent receives the injected content, should this really be considered an unsuccessful detection? I understand the authors' concern about premature intervention, but treating all early detections as failures seems overly restrictive.

I also have concerns about the fairness of the comparisons between content-framed and action-framed monitors. These monitors are designed to identify different signals. A content-framed monitor may correctly identify a malicious instruction even when the agent subsequently refuses to follow it, whereas an action-framed monitor focuses on whether the agent performs a harmful action. Consequently, the poor AUC of content-framed monitors on prompt injection does not necessarily indicate that they fail to detect the injection itself. Rather, it may reflect a mismatch between their detection objective and the benchmark labels. Could the authors more clearly distinguish attack-attempt detection from successful-misuse detection? A comparison that separately evaluates these two objectives would help establish whether action framing genuinely improves monitoring capability.

The experimental results also appear sensitive to the evaluation protocol. For example, Table 12 shows that restricting the sequential monitor to assistant-message positions increases its interval credit from 0.166 to 0.410, approaching the 0.433 achieved by sequential_action. This suggests that much of the reported improvement comes from the monitoring cadence rather than better detection capability.

Finally, I appreciate the detailed reproducibility statement and the authors' intention to release the benchmark. However, the publicly linked GitHub repository currently contains only a README.

2 comments
AuthorsOpenReview ↗

Response to Reviewer EgSB (1/2): real executions, a second generator, and the harm window

We thank the reviewer for a careful review and concrete suggestions. We ran each analysis the reviewer asks for and added them to the revised paper as a new Appendix H. Most results support the paper's conclusions and two lead us to narrow claims.

Validity of Synthetic Traces

We evaluated the monitors on 1,592 real AgentDojo executions (important_instructions attack against 21 agent models) with 794 successful and 798 failed injections (App. H.1). Our main findings still hold. We find that content-framed monitors cannot separate successful from failed injections (Gemma sequential AUC 0.423) while they do detect the attempt. We also find that action-framing increases AUC by +0.484 [0.457, 0.511] on Gemma (+0.443 [0.414, 0.471] on Qwen3.6), and performs better on all 21 agents.

Real runs are indeed harder (Gemma sequential_action AUC score drops from 0.985 -> 0.908) and recall at 10% FPR also drops (0.997 -> 0.548 for Gemma; 0.999 -> 0.785 for Qwen3.6). As §3 notes, a benchmark is easier than the real world, so its results likely overestimate real-world safety. §6 now summarises the real-run results and App. H.1 gives them in full. The direction of our findings is unchanged.

Additional runs from a secondary model

The submitted version already tests the goal leak directly: re-completing 50 traces without the goal moves sequential prefix AUC from 0.885 to 0.881 (App. C.5.1). We further regenerated a subset with Qwen3.6 as a second generator (App. H.2). For decomposition we used an abliterated version of the model to avoid the need to tune the prompt to bypass refusals. We observe that each arm's AUC score moves by −0.012 to +0.020 on injection and by +0.002 to +0.030 on decomposition, thus the conclusions still hold.

We also checked whether a monitor favours traces written by its own model. Gemma and Qwen3.6 each monitored traces from both generators. The Qwen3.6 monitor’s AUC advantage on Qwen-written traces, minus the Gemma monitor’s advantage on those traces, is +0.005 on injection and +0.009 on decomposition (most of the small decomposition value comes from one arm, eae). So neither monitor favours its own model’s traces in a meaningful way.

On the Definition of the Harm Window

We agree that a flag on the injected content can be useful. In Table 9 of Appendix D.7, we had shown the results of a grace-period analysis where we widen the injection window to credit a flag on the tool output carrying the payload, which is the case the reviewer describes. In the revised version, Appendix H.3 extends this to nine different credit rules. They include full credit for flagging the injected tool output (your example) and partial credit λ for any flag before execution.

On injection the conclusion does not depend on the rule:

  • On your example rule: content-framed arms reach at most 0.10 at FPR ≤ 0.1, against at least 0.64 for action-framed arms.
  • Under all nine rules: the advantage of action over content-framed arms stays at least +0.47 at the lower 95% bound.

Content-framed arms rarely flag the injection itself in time. At most 0.6% of their flagged traces are flagged before the payload, and 53–65% only after the harmful call has executed.

On decomposition we concede part of the point. The gap between recall and interval credit does depend on giving early flags no credit. At λ = 0.5 every arm stays at least 0.21 below its recall. Under full credit for early flagging, the metric falls to within 0.035 of recall. However, an action-framed arm leads for every λ≤0.8\lambda \leq 0.8 in all bootstrap replicates. At least 97.8% of action-framed arms' early flags come before the request that elicits the first harmful commitment, a median of 7–9 messages earlier.

§3 now presents interval credit as the strictest of a family of credit rules, and App. H.3 shows recall as the most lenient.

AuthorsOpenReview ↗

Response to Reviewer EgSB (2/2): attempt vs success detection, cadence, and the artifact

Detecting attack attempts vs successful attacks

We agree that the low AUC of content-framed monitors on injection reflects a mismatch between their objective and our labels, not a failure to see the injection. The submitted version notes that every injection trace (including the benign ones) carries the instruction, so content monitors score it either way (App. D.7, Figure 12). But it did not measure attempt detection separately. We now evaluate the two objectives separately on real AgentDojo runs (App. H.4). This includes 904 traces of normal agent workflows with no attacks (also taken from AgentDojo).

  • Content-framed monitors detect the attempt well: clean vs resisted AUC is 0.979 for sequential. But they cannot differentiate between runs where the model resists from those where the model follows the attack and commits harm (0.43–0.49).
  • Action-framed monitors separate resisted from succeeded runs (0.908 for sequential_action). As intended, they barely react to an attack the agent resisted (clean vs resisted 0.59), since no harmful action occurs.

The two framings are answering different questions. Content-framing asks whether an attack arrived, so it also flags unsuccessful attacks. Action-framing asks whether the agent acted on it (and thus caused harm). That is the event a monitor must catch to prevent harm, and is common to any type of attack, and so this is what we have built our benchmark around. This result used the position-blind AUC scores and detection here simply means detecting AgentDojo's explicit <INFORMATION> template; we make no claims about stealthier injections.

§5 now says the two framings answer different questions and no longer calls content-framed monitors coin flips.

Is the gain from framing or from where the monitor scores? (Table 12)

You are right for decomposition, and the submitted paper reaches the same conclusion (App. E.1, §6). There, both framings catch almost the same traces. The timing advantage of sequential_action comes mostly from scoring only at the agent’s messages.

In the original submission we had not run this control on injection, which our claim relies on more. We have now added this (App. H.5). If we score sequential only at the agent’s messages, its AUC on Gemma rises from 0.521 to 0.646, still far below 0.985 for sequential_action. If we further add one sentence to its prompt that asks it to judge the agent’s own actions, this closes most of the remaining gap (86% on Gemma, 91% on Qwen3.6).

The two effects are thus separate. On decomposition, where the monitor scores mostly changes timing. On injection, what it is asked to judge mostly changes detection. Action-framing is still a method that allows one monitor to do well in both settings. §5 now states this more explicitly.

On the Incomplete Artifact

The full code and dataset were provided in the supplementary material submitted with the paper. This also contains per-message monitor scores and scripts to recompute all the tables without a GPU. We deliberately held back the public GitHub release because the decomposition traces contain step-by-step harmful agent behaviour. On publication, we will release the code on GitHub, including the pipeline that sources the seeds and generates the traces. It will not include the pre-computed source seeds or the final traces. The dataset will be released on Hugging Face with gated access to prevent misuse.

All revisions are listed in our general response, and revised text is highlighted in blue.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reviews. We have uploaded a revised PDF and updated supplementary material. Revised text is highlighted in blue. The main changes are listed below.

New Appendix H: additional analyses

  • H.1 Real AgentDojo executions. We scored the monitors on 1,592 real runs from 21 agent models, where the agents produced their own outcomes. Action-framing still raises AUC by 0.44–0.48 on both monitors. (EgSB, 6vEf, ZyLs)
  • H.2 A second generator. We regenerated a subset with Qwen3.6. Each arm's AUC moves by at most 0.03, and neither monitor favours traces written by its own model. (EgSB, 6vEf)
  • H.3 Alternative credit rules. We re-scored every arm under nine credit rules, including partial credit for early flags and full credit for a flag on the injected payload. (EgSB, ZyLs)
  • H.4 Attempt and success. Using 904 clean AgentDojo runs as negatives, we measure detection of the attack attempt and of its success separately. (EgSB)
  • H.5 Matched cadence on injection. This separates where the monitor scores from what it is asked to judge. (EgSB, ZyLs)
  • H.6 Evidence-onset labels. One author labelled 40 traces blind. We also re-scored every arm under five alternative labellings. (6vEf, ZyLs)
  • H.7 Narration stripping and concealed attacks. (ZyLs)
  • H.8 Natural refusals. We compare the refusal twins with refusals from an agent that was never told to refuse. (6vEf, ZyLs)

Main text

  • Abstract and §5. Content-framed monitors detect the injected instruction but not whether the agent follows it, so we no longer call them coin flips. §5 now separates the effect of where the monitor scores from the effect of what it is asked to judge.
  • §2. We cite Brown et al. (2026) and contrast time-to-detect with the harm window.
  • §3. Interval credit is presented as the strictest of a family of credit rules.
  • §4. It now points to the checks on the generator, the labels and the real runs.
  • §6. It reports the real-run results, the second generator, the human label check and the partial-credit results. A flag on a refusal twin is now described as using audit budget rather than as a false positive, since whether it is a cost depends on the deployer's policy. The ethics statement says the same.

Refusal twins rebuilt

We rebuilt all refusal twins on the final evidence-onset labels (1,536 twins). §6, App. C.2, D.11 and E.2 report the updated numbers, and the conclusions are unchanged. (ZyLs)

Other changes

  • App. F now cites STING. (6vEf)
  • App. E.1 clarifies the early-flag inspection: 18 of 21 early flags fall on preparation. (ZyLs)
  • Release plan. The reproducibility statement now gives the release plan. Code will go on GitHub without the precomputed seeds or final traces, and the dataset will go on Hugging Face with gated access. The supplementary material now also includes the code, data and scores for every App. H analysis. (EgSB)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.