acceptodds
Under review as a conference paper at ICLR 2027

When and Why Do Distributed Attacks Evade Code Monitors?

Abstract

When an untrusted coding agent works on a repository with a series of pull requests (PRs), work for a hidden side task can be distributed across PRs. Spreading the attack may or may not help against a monitor that reviews each PR's diff, depending on the side task. We define criteria for decomposable side tasks that in theory allow an attacker to do every component more innocently. We build a dataset of decomposable side tasks. Across more than 11,000 trajectories from four attacker models and eight monitor models, gradual attacks are generally more evasive than the best single-PR placement on the cache-poisoning side tasks (43 of 252 comparisons are significant after Holm correction, none in the other direction), while single-PR attacks tend to be more evasive on the credential-exfiltration side tasks. We study detectability of these attack types both through what the monitor sees and judges, and through activation probing. The most relevant factors are how well the component fits the scope of the current task, if the monitor checks diffs against the task's requirements, how much suspicion honest code raises, and connections of current changes to earlier changes. Linear probes on the attacker's residual stream predict its next unauthorized action about equally well under both strategies. We release the dataset and code.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI control, AI safety, LLM agents, coding agents, distributed attacks, monitoring, persistent state

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HuMbOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies when spreading a hidden side task over several PRs helps an untrusted coding agent evade a per-PR diff monitor. The authors define criteria for decomposable side tasks. They build five multi-tenant Flask services with a cache-poisoning side task. Two credential-exfiltration CLI tasks from Hills et al. serve as the contrast. Over 11,050 trajectories, gradual attacks beat the best single-PR placement on cache poisoning. 43 of 252 comparisons are significant after Holm correction, and none go the other way. On the credential tasks, single-PR attacks appear more evasive. The mechanism analysis points to task scope, requirement checking, the honest-score tail, and definitions missing from the review input. Probes on the attacker reach AUC around 0.87 under both strategies.

Critical strengths

The evaluation protocol is careful. Single-PR attacks are compared at their most evasive placement, chosen per monitor and threshold. False-positive rates are checked on held-out honest runs. Missing outcomes are kept as intervals, not imputed. I found this convincing. It matters because ordering claims between attack strategies are easy to get wrong.

I liked the five-input study in Section 5. Excerpts without the operative rule (R) keep the gap. Excerpts without an unrelated helper (U) close it. This is clean causal evidence for what the reviewer misses. In my view, it is the most useful finding for monitor design. The data, prompts and full monitor database are also released.

Critical weaknesses

I am not entirely sure the “why” part says much beyond the task design. The PR 2 and PR 3 requests say “do NOT edit” the earlier modules (Appendix A.1). A last-position single-PR attacker has to break that sentence. Appendix E.3 confirms that 100% of successful last-position single-PR attacks edit an earlier module, against 0% of honest PRs. These edits provide an explicit out-of-scope cue for the monitor. The “best placement” is also, in practice, the last position. Earlier positions yield very few successes because later honest PRs overwrite the planted modules (Appendix E.4). I therefore cannot tell how much of the gradual advantage is driven by this explicit restriction. The answer “scope fit decides the direction” is then close to the construction itself. I suggest an ablation that removes the “do NOT edit” sentences and keeps everything else.

The “when” part rests on one strong side and one rather weak side. On the credential side, the best-placement comparison in Table 1 covers Expense Tracker only, with 22 successful gradual attacks and 7 successful placement sets; the two-sided Fisher p is 0.55. The second experiment uses a different attacker (GLM-4.5-Air) and monitor (GPT-4o), and has no complete placement sets. Yet Figure 1(b) sits next to Figure 1(a), although the attacker, monitor and calibration protocol differ. This limits what the contrast tells us about the effect of the side-task family. I also wonder how this fits with Hills et al., who find gradual attacks more evasive under the GPT-4o task-description diff monitor, including on triggered-exfil. The cache side is narrow too. All 43 Holm-significant comparisons in the main experiment come from one attacker, Qwen3.8-27B. The Sonnet 5 extension adds one significant comparison, but evidence across attackers remains limited. The gap is also most pronounced under Mistral-24B, GPT-OSS-120B and GLM-4.7-Flash. Without stronger evidence for the credential contrast, the remaining empirical claim overlaps substantially with Hills et al.

Questions for authors

Clarifications:

  1. Table 1 uses thresholds calibrated on 20 honest Expense Tracker trajectories, while Appendix G uses 53 pooled histories from both tools. Why do these analyses use different calibration populations, and how sensitive are the credential results to this choice?

  2. Could the authors explain the difference from Hills et al. on triggered-exfil under the GPT-4o task-description diff monitor? How much comes from the attacker, attack prompt or evaluation protocol?

  3. The gradual advantage is most pronounced under Mistral-24B, GPT-OSS-120B and GLM-4.7-Flash, while Sonnet 5 catches almost all successful attacks of both kinds in the reported cells. How should we reconcile this with Hills et al.’s finding that the gap grows for stronger monitors?

  4. How were the 42 undefined Sol-attacker cells handled in the Holm correction? Were they retained in the family size of 252?

Suggestions:

  1. Does the gradual advantage survive without the “do NOT edit” sentences?

  2. Could the authors run the credential tasks under the main protocol, with the same attacker configurations, 200 honest calibration runs, complete placement sets, and nominal 1% and 2% FPR thresholds?

  3. Could the authors evaluate the Section 8 recommendations on the main attack trajectories? Section 5 tests access to definitions on constructed bugs, and Appendix F tests a link-tracker, but its honest calibration is incomplete. A full comparison could include explicit requirement checking, access to relevant definitions, and stateful monitors such as:

    • link-tracker [Hills et al., COLM 2026]
    • the sequential monitor [Chen et al., ICLR 2026]
    • the hybrid hierarchical-sequential monitor [Kale et al., ICLR 2026]

    These could be evaluated using the saved trajectories, with suitable adaptations and fresh honest calibration, to test whether the gradual advantage survives.

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

This paper investigates when distributing malicious code changes across multiple PRs helps evade LLM-based code monitors. It introduces decomposable attack tasks and evaluates gradual versus single-PR attacks across over 11,000 trajectories. The authors further analyze monitor failures through code-context interventions, qualitative analysis, and activation probing.

Critical strengths
  1. Careful experimental design. I appreciate the comparison against the best single-PR placement, along with behavioral success checks, calibrated detection thresholds, and multiple-testing correction.

  2. Insightful analysis of monitor failures. The controlled study in Section 5 is particularly interesting, showing that providing the relevant code definitions can largely eliminate the detection gap. This offers useful insights for improving agent monitoring.

Critical weaknesses
  1. Limited attack diversity. Most cache-poisoning tasks share the same underlying vulnerability, and all 43 significant comparisons in the main experiment come from the Qwen attacker. The generalizability of the findings remains somewhat unclear.

  2. Task-design dependence. Single-PR attacks necessarily modify earlier modules despite explicit file restrictions, making them easier to detect by construction. It's unclear how much of the advantage comes from distributing the attack itself versus avoiding these obvious scope violations.

Questions for authors

See above.

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

This paper investigates whether splitting a task across multiple pull requests (PRs) evades per-PR monitoring. By formalizing decomposability of tasks, the paper shows that multi-PR attacks are less detectable than single-PR attacks. Empirical analysis shows that evasion improves when staying within each PR's scope.

Critical strengths
  • The four criteria for decomposability are a useful set of rules for future research.
Critical weaknesses
  • The attack design seems to be specific to data exfiltration. Even the prompt in Appendix B focuses on data exfiltration. It is unclear whether this design generalizes to other types of attacks.

  • Design Criterion 3 ("no two [task components] share a variable, import or function name") seems to be violated at least in the example of Appendix A, where all three PRs include an instruction to "Start it with from substrate import app.", thus sharing imports. Maybe the Criterion 3 should be more loosely defined, though this needs to be measured in terms of impact on evasion.

  • Some terms or concepts are not defined before use, or are used too early in the paper.

    • Figure 1 in the introduction uses the phrase "honest-calibrated threshold on the trajectory maximum" without explaining it.
    • The terms "gradual" and "distributed" seem to be used interchangeably when talking about the attacks and the adversaries, though in general they are not synonyms.
    • "behavioral scorer" is discussed in the main body of the paper but never described until Appendix D.
  • The related work lists relevant papers, but does not discuss differences from the present paper. I would like to see more discussion on how each of Lindner et al. (2025), Phuong et al. (2026), Makins et al. (2026), Remedios et al. (2026), and Najt et al. (2026) differ from this paper.

Questions for authors
  • The implications for monitor design in Section 8 mentions "changes in what the monitor sees" and not "a new architecture". Can you be more specific on what changes would mean a new architecture?

  • Can you argue why your criteria for decomposability are applicable to attacks beyond unwanted information flows like exfiltration and cache poisoning?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.