acceptodds
Under review as a conference paper at ICLR 2027

TraceEdit: Self-Evolving Multi-Channel Environment Attacks on GUI Agents

Abstract

GUI agents built on multimodal large language models can be hijacked by adversaries who mutate the agent's environment, and reported persistence of that hijack across planning, memory, and tool-use stages is a recurring empirical finding. We argue that this finding is partly an artifact of measurement: when the adversarial trace is still on the screen during evaluation, observed persistence conflates continued exposure with an internalized shift in the agent's behavior, and the latter remains underexplored as a training objective. Our claim is that persistence should be defined and trained counterfactually, by removing the adversarial trace before re-evaluating the agent. We operationalize this claim with TraceEdit, a self-evolving attacker that maintains a reversible edit trace over five environment injection surfaces (grouped into channels: text, visual saliency, layout, interaction flow, and event timing) and is trained under a constrained reinforcement-learning objective whose persistence term is computed by a deterministic counterfactual re-evaluation in which every reversible edit has been undone. Multi-channel coordination is the setting in which rollback-conditioned persistence has the most room to differ from exposure, and a dual variable enforces a fixed detectability constraint on every emitted edit. Empirical results across four GUI-agent targets and six security benchmarks demonstrate that TraceEdit outperforms a broad set of baselines on attack effectiveness, residual persistence, and stealth, and that its rollback-conditioned persistence is empirically separable from on-screen-time.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
GUI Agents, multimodal large language models, reinforcement learning, agent security

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_xYM5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes TraceEdit, a self-evolving attack framework for GUI agents. It coordinates five environment-editing channels and introduces a counterfactual persistence reward that measures whether malicious influence remains after adversarial edits are removed. The attacker is trained with constrained reinforcement learning to balance attack success, persistence, and stealth. Experiments across six benchmarks and four GUI agents show improvements over thirteen baselines, supported by ablations and robustness analyses.

Critical strengths

1 The paper addresses an important measurement problem: stopping edits does not necessarily stop exposure to an attack. 2 The appendix provides substantial controls, including full-DSL versions of prior attackers, channel-by-reward ablations, environment restoration, and replay checks. 3 Evaluation extends beyond attack success to human detectability, judge agreement, transfer, and defenses, offering several useful perspectives.

Critical weaknesses
  1. I have a major concern regarding the counterfactual rollback protocol. Section 3.2 applies Undo immediately after the last edit, while Table 7 states that pending timers are canceled. However, in Tables 36–37, scheduled navigation occurs at step 6, after the rollback point at step 5, yet the examples still report RHLremoved=3. Could the authors clarify how these navigations survive rollback? This seems important for the validity of the central persistence metric.

2 The human stealth results do not agree across presentations. Figure 12 labels its vertical axis 1-HDR, with TraceEdit’s benchmark points around 0.83–0.91. These imply HDR values around 9–17%, whereas Tables 1 and 32 report 27.9% under the same stated setting. Are different subsets or annotation units being used? This matters because stealth is a central claim.

Questions for authors

I find this to be a very solid and carefully executed work. I would appreciate the authors' responses to my questions above, and I am open to raising my score if these concerns are adequately addressed.

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

This paper introduces TraceEdit, a reinforcement-learning attacker that coordinates GUI-environment edits across text, visual saliency, layout, interaction flow, and event timing. Its central contribution is a rollback-conditioned persistence objective: after the attacker stops editing, the evaluator reverses reversible edits while preserving the agent’s history and measures subsequent actions advancing the malicious goal. The attacker combines a grammar-constrained edit language, per-reward advantage normalization, and a Lagrangian naturalness constraint. Experiments cover multiple GUI agents and six adapted security benchmarks, with comparisons against prior attackers, same-action-space search baselines, channel-by-reward ablations, replay tests, human detectability assessments, and additional environment-restoration controls.

Critical strengths

The paper identifies a meaningful measurement distinction between continued exposure to an adversarial artifact and influence surviving its removal. The reversible edit records make this distinction operational and allow persistence to become a training objective. The additional environment-restored continuation in Table 28 is particularly useful because it separates consequences carried by application state from those associated with the agent’s retained history. The evaluation includes substantial attempts to isolate the proposed components. The channel-by-reward factorial distinguishes the effects of broader editing access and the counterfactual reward, while same-DSL search baselines and the expanded prior-attacker comparisons in Table 16 address some action-space confounds. Replay stability, human assessments, and judge-reliability checks also provide useful evidence beyond aggregate attack success rates.

Critical weaknesses

The persistence metric does not yet cleanly support the broader interpretation of persistent agent hijacking. Appendix A.4 averages residual hijack length only over successful attacks, so comparisons can involve different selected task subsets. The unnormalized count also depends on the remaining continuation length, and the policy controls when its final edit occurs. Furthermore, rollback retains non-invertible effects and agent-committed application state. The restoration control acknowledges this issue: TraceEdit’s residual decreases from 1.7 to 0.9 steps after environment restoration, with 0.5 steps above the clean background attributed to history-conditioned influence. These controls are valuable, but the central reward factorial is reported on the broader rollback endpoint. Paired, horizon-controlled results on the restored endpoint are needed to establish how much of the persistence gain reflects influence retained by the agent rather than selection, timing, or surviving environment changes. Several reporting and implementation ambiguities prevent a confident assessment of the experimental evidence. In Table 13, the “Combined (any)” non-invertible-trajectory fraction is lower than the TimingEdit fraction in every benchmark; this is impossible under a common trajectory denominator and requires explicit denominator definitions or corrected results. Appendix B.5 and Table 30 also describe Claude Computer-Use variants fine-tuned with memory-hygiene and provenance auxiliary losses, without specifying the trainable checkpoint or implementation enabling these updates. Finally, the abstract describes a constraint enforced on every emitted edit, whereas the optimization section states enforcement in expectation with a positive slack tolerance. These issues affect the rollback-validity analysis, defense experiments, and interpretation of the stealth comparison.

Questions for authors
  1. What are the exact denominators for each row of Table 13? How can the fraction of trajectories containing any non-invertible edit be lower than the TimingEdit fraction? Please provide the corresponding counts and explain whether correcting this changes Table 14 or the main persistence results.
  2. What model or component is actually trained in the “defense-aware-trained Claude Computer-Use” experiments? Please specify the checkpoint, training interface, auxiliary-loss implementation, and whether the evaluated system remains the stated Claude target.
  3. Does the counterfactual reward improve persistence on a common paired task set, with a fixed post-edit evaluation window and application state fully restored? Please report the reward factorial under these conditions, including uncertainty for the claimed positive interaction.
  4. Is the naturalness threshold enforced by rejecting individual edits, or only through an expected penalty? Please report violation rates before any filtering and distinguish the learned policy’s feasibility from feasibility imposed by the evaluation harness.
Reviewer_RY1dOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TraceEdit, a learned attacker that coordinates environment edits across five channels: text, visual saliency, layout, interaction flow, and event timing. Its central proposal is to evaluate and optimize attack persistence after reversible adversarial edits have been removed, distinguishing residual influence from continued exposure to malicious content. The method combines a structured editing language, reversible edit records, a counterfactual persistence reward, and constrained reinforcement learning. Experiments across six benchmarks and four GUI-agent targets report improvements in attack success, residual persistence, and stealth.

Critical strengths
  1. The paper correctly highlights that continued malicious behavior while an adversarial artifact remains visible is insufficient evidence of influence surviving artifact removal.
  2. The paper proposes a unified framework for coordinated environment attacks. The five-channel editing language enables combinations of textual, visual, spatial, interactive, and temporal manipulations within one policy.
  3. The evaluation covers multiple targets and benchmarks and includes stealth assessment, defense experiments, and rollback-validity checks.
Critical weaknesses
  1. The absolute improvements over strong baselines are modest, particularly for persistence. The practical significance of these gains is unclear.
  2. Limited evidence of durable compromise. Post-removal effects average only 1.4–2.1 malicious actions, providing limited support for prolonged hijacking.
  3. Persistence relies on an LLM judge, and reported false positives raise concerns given the small margins over strong baselines.
Questions for authors
  1. Could you please report confidence intervals for persistence results?
  2. Can residual influence be decomposed more explicitly into contributions from contaminated history, agent-committed application state, and incomplete rollback?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.