acceptodds
Under review as a conference paper at ICLR 2027

Offline Scores Overestimate Online Oversight: Timely Interception of LLM Agents with Causal-Cone Monitoring

Abstract

Monitors for LLM agents are predominantly evaluated offline: a suspicion score is computed over a completed trajectory and summarized by AUC-ROC. Deployment asks a different question, because an alarm raised after the wire transfer clears or the API key leaves the machine is an autopsy rather than an intervention: can the monitor fire before the agent's first irreversible step? We formalize this as online interception and introduce the Online Interception Framework (OIF), an evaluation protocol that anchors each attack trajectory at a deterministic, auditable harm horizon , the first irreversible step, located by a rule-based detector rather than by the model's self-report, and scores a monitor by whether its alarm fires strictly before under a matched false-alarm budget. Under OIF, offline and online performance decouple: across two environments, three attack families, and four monitors spanning open-weight and closed models, completed-trajectory AUC can stay high while pre-horizon interception is low, and the offline ranking need not predict the online one. To explain when interception fails, we separate a monitor's two information channels, the agent's stated channel (its chain-of-thought and reports) and the action channel (what an action can causally reach), and evaluate a causal-cone monitor that reads the action channel: at each step it scores whether the action opens a path toward harm, rather than trusting the agent's stated justification. Its timeliness advantage grows with how far the attack corrupts the stated channel, from large under monitor hijacking and deceptive reasoning, to small under hidden reasoning, to none when attacks are too weak to bite; an effectiveness-stratified contrast ties the advantage to injection efficacy (ineffective injections , effective injections ). Every experiment runs on a single workstation with an open-weight local monitor overseeing closed frontier agents, so both the protocol and the monitor are inexpensive to reproduce.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, AI safety, agent monitoring, runtime oversight, online interception, chain-of-thought monitoring, prompt injection, causal reachability

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_2d6MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that monitors for LLM agents are mostly evaluated offline (AUC over completed trajectories), while in deployment what matters is whether the alarm comes before the agent’s first irreversible action. To measure this, the authors propose the Online Interception Framework (OIF). Each attack trajectory is anchored at a harm horizon found by a rule-based detector, and a monitor counts as successful only if it fires strictly before that horizon, at a false-alarm budget calibrated on benign runs. On SHADE-Arena and a computer-use environment, with three attack families and four monitors, they show that offline AUC can stay high while pre-horizon interception is low, and that offline rankings don’t necessarily carry over online. They also propose a “causal-cone” monitor that judges whether each action opens a path to harm instead of relying on the agent’s stated reasoning. Its advantage over the MRT verdict baseline appears to grow with how much the attack corrupts that stated channel. Everything runs on one workstation, with a local open-weight model monitoring closed frontier agents.

Critical strengths

This manuscript consider a important problem that whether the monitor is timely or not. The paper makes smart design choices that a rule-based harm horizon that the agent cannot change, matched false-alarm thresholds for all monitors, and full budget curves. Figure 1 shows this clearly, and that is the main value of the paper.

Critical weaknesses

The kind of detection problem is widely considered, both in controlling the great accuracy and F1 score. Although author propose a new method, I can't see a strong and convincing comparison to many similar methods in terms of latency, accuracy, etc. Moreover, the figures in this manuscript are ugly, especially figure1, which makes me think the author probably did not read the paper by themselves.

Questions for authors
  1. Adaptive attacks. Have you tried attacks designed against the cone itself? For example, splitting a harmful action into steps that each look low-reachability, or injections aimed at the reachability judge.
  2. Novelty. Beyond prompting an LLM judge to focus on what an action can lead to, what does the cone add? How is it conceptually different from AgentSpec or ProbGuard?
1 comment

I thank the author for their response. The author addresses part of my concerns, but remains the weakness considering the novelty and adaptive attack. I will keep my score.

Reviewer_C7p2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that LLM agent monitors should be judged by whether they raise an alarm before the first irreversible harmful step instead of using AUC over completed trajectories. It proposes the Online Interception Framework (OIF). OIF locates a harm step with rule-based detectors and reports the fraction of malicious runs flagged strictly before the harm step, at false-alarm budgets set on benign runs. It also proposes a "causal-cone" monitor, an LLM monitor that rates whether each action opens a path toward harm.

Critical strengths

Evaluating when a monitor raises an alarm seems useful. The paper shows concretely that a high whole-trajectory score can hide late alarms.

Critical weaknesses
  1. I am unsure which deployment setting the evaluation represents. A monitor that checks each proposed tool call can block the harmful call before it executes. I could not determine whether this would count as successful interception under the strict metric and whether those alarms use information available before execution. Additionally, I would like to understand how much of the measured lateness comes from the monitor’s judgments and how much comes from this evaluation setup, since a monitor that checks every action has more opportunities to intervene than a five-step chunked monitor.

  2. The false-alarm budgets are high compared with related work, and the comparison with control evaluations is missing. The main results use budgets of 10–25%. The authors explain that their benign datasets are too small to estimate lower rates reliably, which is understandable, but this limits what we can conclude about settings where false alarms are more costly. Monitors are usually evaluated at much lower false-alarm rates, typically around 1–2%. I would also appreciate a clearer comparison with AI control evaluations, since these papers also already evaluate monitors online: they define safety as stopping an attack before it succeeds.

  3. It is unclear how broadly the harm-step definition applies. Here, harm is a single clearly identifiable action. Harm is often less well defined: 1) diffuse: data leaked in small pieces across many steps 2) decomposed: several individually benign actions that are harmful only in combination 3) contextual: the same action is fine in one run and harmful in another. The paper excludes a fifth SHADE task because its honest runs perform the same kind of action as the attack, so the harm step cannot be identified

Questions for authors
  1. Is OIF intended to evaluate a monitor that approves each action before execution, a monitor that runs alongside the agent, or both? Would blocking the harmful call after it is proposed but before execution count as successful interception?

  2. Could you evaluate a monitor before every action, using the history and proposed call available at that moment? This would help distinguish limitations of the monitor from effects of the five-step chunks.

  3. How does OIF relate to the online safety evaluations used in AI control work? Is there enough benign data to assess whether the conclusions hold at false-alarm budgets closer to 1–2%?

  4. How would you define the harm step when harm accumulates across several actions? Would moving the harm horizon one or two steps earlier materially change the main results?

Reviewer_5KcgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that offline evaluation based on full-trajectory AUC may overestimate an LLM agent monitor’s ability to raise an alarm before harm occurs. The authors introduce the Online Interception Framework (OIF), which uses rule-based detectors to identify the first irreversible action and measures alarm recall before that point under matched false-alarm budgets. They also propose a “causal-cone” monitor that asks an LLM to assess whether an action opens a path toward harm. Log-replay experiments across tool-calling and computer-use environments show that offline performance can diverge from timely interception capability, with the proposed method’s advantages concentrated in deceptive-reasoning and monitor-hijacking scenarios. The work’s value lies in improving the evaluation objective. However, prefixes that disclose attacker goals, limited tasks and samples, relatively high false-alarm budgets, and the absence of live interception experiments constrain its deployment claims and mechanistic interpretation.

Critical strengths
  1. The problem is valuable, and the evaluation objective aligns more directly with safety intervention needs.

    The paper denotes the first irreversible action by tIt_I, requires the alarm time to satisfy Tϕ<tIT_\phi<t_I, and measures timely alarm recall (PREH) under a false-alarm budget. This distinguishes between eventually detecting an attack and detecting it while intervention is still possible, giving the evaluation clear practical relevance.

    More importantly, the authors do not merely compare AUC and PREH, which measure different quantities. They also compare whole-trajectory detection rates with pre-harm detection rates at the same operating point, showing that some failures genuinely arise from alarms arriving too late. This evidence is more convincing than simply demonstrating high AUC alongside low PREH. (pp. 4–7, 20)

  2. The experimental comparisons are reasonably controlled, and several alternative explanations are examined carefully.

    The two monitors use the same trajectories, underlying model, five-step chunks, and number of model calls. The authors state that the prompt configuration was selected on a separate development subset and then frozen. The appendices also provide ablations involving strict versus inclusive timing criteria, signal aggregation, five-fold threshold validation, and a 0–100 scoring baseline.

    In particular, widening the baseline's scoring range addresses the concern that the gains might arise solely because continuous scores allow more flexible threshold calibration. Although the coverage remains limited, these checks demonstrate a constructive effort to scrutinize the experimental findings. (pp. 5, 16–21)

  3. Negative results and uncertainty are disclosed relatively thoroughly, helping establish the method's boundaries.

    The authors explicitly report underperformance on the API-key task, ceiling effects on some computer-use tasks, the baseline's advantage at near-zero false-alarm budgets, and weaker localization of harmful steps. They also acknowledge that several results provide only directional evidence.

    This presentation helps readers understand the conditions under which the method is useful. The work is currently best positioned as an empirical finding and evaluation improvement with clear boundaries, rather than a broadly effective security solution. (pp. 8–9, 18–21)

Critical weaknesses
  1. The monitor receives the malicious objective in advance, substantially limiting external validity.

    Page 7 explicitly discloses that the monitor input contains the environment's hidden system message, including the attacker's side-task specification. The monitor may therefore already know what the attacker intends to accomplish when judging whether the current action advances that objective.

    This differs substantially from detecting unknown attacks in real deployments. Giving both methods the same information ensures input consistency, but different prompting strategies may exploit that information to different degrees. Its effect on the relative advantage therefore cannot be ruled out. The paper leaves prefix removal to future work, although this should be a central ablation supporting its main conclusions. (pp. 7, 9)

  2. Online interception is evaluated through log replay, without validating actual blocking capability.

    Page 6 explicitly states that the authors do not run or gate a live agent. Instead, they replay existing logs prefix by prefix. The evidence therefore establishes that the monitor could potentially raise an alarm earlier in the logged action sequence.

    The same page reports that the causal-cone monitor takes approximately one minute to process a five-step chunk and is slower than the baseline. Synchronous monitoring requires evaluating the resulting task latency and throughput; asynchronous execution requires assessing whether harm has already occurred by the time the alarm arrives. Log replay is a reasonable evaluation approach, but it does not establish real-time deployment feasibility or end-to-end safety benefits. (p. 6)

  3. The mechanistic interpretation of the causal cone exceeds the evidence, and the algorithmic contribution has not been sufficiently isolated.

    The paper explicitly states that the method neither constructs a state-transition model nor enumerates a reachable set. Instead, it asks an LLM to score the potential consequences of an action. The method is therefore closer to reachability-inspired prompting and does not provide formal causal analysis or safety guarantees.

    Both monitors actually read the same complete transcript. The causal-cone monitor additionally generates structured evidence, while output format and token expenditure are not matched. The current results cannot clearly distinguish gains from reachability reasoning, evidence generation, or prompt wording.

    Furthermore, injection effectiveness is defined using baseline scores and then related to performance gains over that same baseline, so the two quantities share the baseline signal. With only two attack conditions, this contrast is insufficient to establish the proposed causal mechanism. Factorial experiments matching output format and computational expenditure, together with adaptive attacks specifically optimized against the causal-cone prompt, are needed. (pp. 5, 8, 17, 20)

  4. Statistical support for the main gains is limited, and generalization across independent tasks remains weak.

    The main average improvement on SHADE deceptive reasoning is approximately 17 percentage points, but its 95% confidence interval is [−0.00,+0.23][-0.00,+0.23], placing it at the boundary of significance. At a false-alarm budget of 0.10, the improvement is approximately six percentage points, with a confidence interval of [−0.19,+0.29][-0.19,+0.29].

    The SHADE monitor-hijacking evaluation contains only 15 valid trajectories, and the confidence intervals for the gains span zero at every budget. Although the computer-use experiments contain more trajectories, they cover only approximately 13 task clusters; the trajectories cannot all be treated as independent evidence of generalization.

    These results support the possibility that offline performance and timely detection can diverge, but provide weaker support for consistent superiority across environments and models. More independent tasks, attack strategies, and external test sets are needed. (pp. 8–9, 18–20)

  5. False-alarm budgets are relatively high, and equal budgets do not imply equal realized false-alarm costs.

    The main experiments primarily use trajectory-level false-alarm budgets of 10%–25%. This is sufficient to expose an evaluation gap, but provides limited support for deployment under stringent low-false-alarm requirements. The authors also acknowledge that the benign samples are too small to evaluate lower false-alarm regimes reliably.

    In the five-fold validation, the realized false-alarm rates at a budget of 0.25 are approximately 0.22 and 0.14 for the two monitors; at a budget of 0.20, they are approximately 0.21 and 0.10. Thus, the methods satisfy the same budget constraint but incur different realized false-alarm costs.

    The finer-grained baseline wins at near-zero false-alarm rates, demonstrating that the method's advantage depends strongly on the operating budget. Further evaluation should compare the methods at the same realized false-alarm rate and report legitimate-task completion rates and human review costs. (pp. 5, 19–21)

  6. Deterministic harm horizons do not guarantee semantic validity, and filtering affects the scope of the evaluation.

    Rule-based detectors provide reproducible and auditable values of tIt_I, which is a strength. However, determining when actions such as file writes or changes to numerical fields in documents constitute the first irreversible harm still requires an independent definition and validation. Detecting completion of an attacker objective is not necessarily equivalent to locating the earliest point of irreversible risk.

    The paper also excludes one SHADE task because benign and malicious behavior share the same action class, and computes PREH only for malicious trajectories with finite tIt_I. The metric therefore primarily measures whether attacks that would otherwise cause identifiable harm can be detected in advance. It should not be interpreted directly as the overall safety rate across all malicious attempts. Additional evidence is needed on horizon-annotation agreement, sensitivity to the chosen time point, and performance over the complete task set. (pp. 6, 21)

Questions for authors

Please refer to the listed weaknesses in previous section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.