acceptodds
Under review as a conference paper at ICLR 2027

What Do Attack-Success Rates Measure? A Matched-Sham Control for GUI-Agent Injection Benchmarks

Abstract

Prompt-injection evaluations of graphical user-interface (GUI) agents often count whether an agent acts on an injected instruction. Such decision compliance is distinct from completing an attacker’s objective, and a malicious-only score cannot show how often an agent would act on a benign instruction delivered in the same way. We use a three-arm comparison—malicious, matched sham, and clean—to measure this distinction under a fixed execution scaffold. On 57 live VPI-Bench browser-use cases, sham instructions were followed in 52/114 first post-exposure actions and 84/114 episodes at any step, compared with 112/114 and 114/114 for malicious instructions; the clean arm was 0/114. In a preregistered page-realism-by-interactivity analysis, sham engagement was 25/80 on toy pages under one-shot probing, 24/740 on toy pages under interactive probing, and 0/119 and 0/120 in the two real-page cells. In email and messenger tasks, the malicious-versus-sham contrast was inconclusive after symmetric rescoring. These results describe decision compliance under our tested scaffold; their rate differences and ratios do not decompose the task-completion rates reported by the original benchmark.

Primary area:
interpretability and explainable AI
Keywords:
Visual agents, adversarial observations, trajectory attribution, root-cause localization, counterfactual replay, multimodal evidence, causal intervention, task recovery, agent robustness, GUI agents

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_v3dxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the limitations of attack-success metrics in GUI-agent prompt injection evaluation. It proposes a three-condition protocol comparing malicious injections, matched benign sham instructions, and clean inputs to distinguish malicious target engagement from general compliance with external instructions. The authors evaluate this protocol across browser-use, email, messenger, and live-web tasks.

Critical strengths
  • I find the problem relevant to GUI agent security, and I appreciate the attempt to better understand what existing attack-success metrics actually measure.
  • I appreciate the matched-control experiments, statistical analyses, and the authors' careful distinction between instruction engagement and actual attacker-goal completion.
  • I also appreciate that the authors report inconclusive results, examine scoring sensitivity, and document experimental issues discovered during their investigation.
Critical weaknesses
  • Prior work has already established several important limitations of attack-success metrics, including the distinction between intermediate engagement and completed attacker objectives [1,6], security–fidelity ambiguity [2], and counterfactual attribution of agent actions [3]. The present paper offers a concrete matched-sham behavioral control, but its incremental insight beyond these established distinctions is not yet sufficiently demonstrated.
  • The authors preserve the injection format and delivery context, but change the requested target behavior. Malicious and benign requests may differ substantially in task difficulty, authorization, and alignment with the original user task. Therefore, I do not think the malicious-minus-sham difference necessarily isolates the effect of malicious semantics. Following a benign request may also be appropriate behavior rather than a security failure.
  • While Table 1 reports 45.61% first-action sham engagement, Table 2 shows no sham engagement across 239 real-page runs, and Table 4 reports only 1/120 on AssistantBench. The email and messenger findings also become inconclusive under symmetric scoring. Although these experiments use different settings, I think they raise important questions about whether the reported phenomenon is broadly relevant or mainly reflects particular benchmark configurations.
  • I am not convinced that the evaluation demonstrates sufficient practical value. The main experiments measure decision compliance under the authors' own executor scaffold rather than the original benchmark's task-completion ASR. The paper does not demonstrate whether matched-sham evaluation changes model rankings, defense comparisons, or conclusions about actual attacker-goal completion. I think this is a significant limitation for a paper whose main contribution is improving the interpretation of existing security evaluations. The absence of raw execution records for the key 2×2 analysis also limits reproducibility.

[1] VPI-Bench: Visual Prompt Injection Attacks for Computer-Use Agents. ICLR 2026.

[2] Security–Fidelity Tradeoffs: The Hidden Cost of Prompt Injection Defense. ICML 2026.

[3] AttriGuard: Defeating Indirect Prompt Injection in LLM Agents via Causal Attribution of Tool Invocations. USENIX Security 2026.

[4] Who Pays the Price? Stakeholder-Centric Prompt Injection Benchmarking for Real-world Web Agents. arXiv 2026.

[5] LivePI: More Realistic Benchmarking of Agents Against Indirect Prompt Injection. arXiv 2026.

[6] WASP: Benchmarking Web Agent Security Against Prompt Injection Attacks. NeurIPS 2025.

[7] AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents. NeurIPS 2024.

Questions for authors
  • Could the authors clarify what is methodologically new beyond standard matched-control design and related agent security evaluations [2,3]? What conclusions can this protocol support that existing evaluations cannot?
  • How are malicious and sham targets matched in terms of task difficulty, authorization, and alignment with the original user request? Could the authors provide a stratified analysis or a small controlled experiment isolating these factors?
  • Why does sham engagement largely disappear in the real-page and AssistantBench experiments? Could the authors further analyze whether this difference comes from page realism, task clarity, or differences in the underlying evaluation cases?
  • Could the authors demonstrate whether the proposed protocol changes an existing benchmark's model ranking or defense evaluation, ideally using actual attacker-goal completion? Could they also provide the raw execution traces and scoring implementation for the 2×2 experiment?
Reviewer_6i1SOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines whether attack-success rates in GUI-agent prompt injection benchmarks reflect genuine malicious influence or general compliance with external instructions. The authors introduce a matched-sham evaluation framework comparing malicious, benign, and clean conditions. Experiments across multiple benchmarks show substantial benign-instruction compliance in some settings, with results varying by environment and scoring criteria. The paper highlights the importance of distinguishing instruction-following behavior from actual attack success and advocates more rigorous evaluation practices.

Critical strengths
  1. Well-motivated research problem. The paper addresses an important and often overlooked limitation of existing prompt injection benchmarks: whether the reported attack success rates truly reflect successful malicious attacks or merely agents' general tendency to follow external instructions. This distinction is critical for accurately evaluating the security and robustness of GUI agents.
Critical weaknesses
  1. Limited evaluation scope. The experiments rely on only two agent models and a relatively limited set of benchmark tasks, raising concerns about the generalizability of the findings across different models, environments, and attack scenarios.

  2. Insufficient discussion of related work. The paper overlooks closely related studies, including [1] The Blind Spot of Agent Safety: How Benign User Instructions Expose Critical Vulnerabilities in Computer-Use Agents and [2] When Benign Inputs Lead to Severe Harms: Eliciting Unsafe Unintended Behaviors of Computer-Use Agents. Both investigate safety risks arising from benign instructions or inputs, which are relevant to the paper's motivation. The lack of discussion makes it difficult to assess the novelty and distinct contributions of this work.

Questions for authors

N/A

Reviewer_8KMcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the interpretation of prompt-injection results for GUI agents by comparing malicious instructions, matched benign instructions, and a clean condition. It evaluates target engagement on VPI-Bench cases and AssistantBench tasks using two executors. The results show substantial engagement with benign instructions in some settings, while the corresponding effect is much smaller in others. The paper argues that matched controls can help distinguish different behaviors underlying reported attack outcomes.

Critical strengths

Comparing malicious, benign, and clean conditions is a useful way to examine what an evaluation metric measures. The paper also explicitly distinguishes target engagement from completed attacker objectives, which is important for interpreting the results.

Critical weaknesses
  1. The practical significance for contemporary agents is insufficiently established. The study uses two incompletely specified executors under its own execution setup. It does not show whether the conclusions persist for representative strong general-purpose agents with explicit threat-awareness instructions and their ordinary safeguards. Moreover, the benign-control effect varies substantially across settings: AssistantBench has only 1/120 sham engagements compared with 52/120 malicious engagements. The paper therefore does not establish how broadly its main observation changes current assessments of agent vulnerability.
  2. The evidence does not establish how the proposed control changes conclusions about successful attacks. Target engagement does not necessarily imply completed harm, and benign engagement may be consistent with the user's task. The executed-arrival comparison is additionally confounded by unequal accessibility of malicious and benign destinations. The toy-versus-real-page comparison also uses different task sets, limiting conclusions about the effect of realism. Although the paper acknowledges these limitations, the remaining contribution is a narrow descriptive finding rather than a well-established correction to attack-success evaluation.
Questions for authors
  1. Agent specification and applicability: What exact model versions, system instructions, tools, and safeguards were used for each executor? Do the matched-control findings persist for a strong general-purpose agent explicitly warned about environmental prompt injection?
  2. Outcome validity: Can the authors evaluate completed attacker objectives alongside legitimate-task completion, with matched destination accessibility? How does adding the sham control change the conclusions under those outcome measures?
  3. Interpretation: How should readers distinguish legitimate compliance with benign task content from susceptibility to unauthorized instructions? What conclusion does the malicious-versus-sham contrast support beyond the separate rates, particularly where sham engagement is nearly absent?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.