acceptodds
Withdrawn from ICLR 2027

ClickShield: A Firewall for Multimodal Computer-use Agents Against Dual Prompt Injection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Multimodal computer-use agents execute goals by repeatedly interpreting a screenshot and a natural-language instruction. Because today’s agents implicitly trust both the user prompt and on-screen content, they are vulnerable to dual prompt injection: malicious instructions and deceptive UI elements (e.g., phishing pop-ups) that steer the agent into unsafe actions. We introduce ClickShield, a training-free, model-agnostic runtime firewall that wraps an existing agent without retraining or modifying its policy. ClickShield enforces a zero-trust boundary via two parallel modules: (i) an instruction risk filter using an LLM-as-a-judge, and (ii) a decoupled UI parser that detects interactable regions, extracts OCR text, and produces lightweight element captions to form a structured, DOM-like screen representation. A judge LLM then compares the verified goal against this element list to flag intent-conflicting elements; ClickShield deterministically blackout-masks their bounding boxes before the agent acts. Across RiOSWorld-Bench (492 risky tasks) and OS-Harm-Bench, ClickShield sharply suppresses attack success (e.g., adversarial goal completion (59.6% 5.4%) while preserving benign utility on OSWorld (only marginal success degradation).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Computer-use agents, Multimodal agents, GUI agents, Prompt injection, Visual prompt injection, Agentic security, Runtime guardrails, Screen sanitization, Vision-language models, Tool use

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

This paper introduces ClickShield, a runtime firewall for multimodal computer-use agents against textual and visual prompt injections. The framework consists of two branches. Textual branch uses an LLM judge to identify risky instructions. Visual branch uses YOLO, OCR, and BLIP to construct a structured representation, followed by another LLM judge. The paper evaluates on RiOSWorld, OS-Harm, and CUAHarm across several models. The paper also reports low inference overhead and relatively small changes in benign task success.

Critical strengths
  1. Important and practical problem. Malicious instructions can be embedded in both user inputs and visual environments, and agents may act on these instructions without recognizing the associated risks.
  2. Simple design. Two branches with LLM judges.
  3. Broad evaluation. The experiments cover 10 multimodal models and several security benchmarks, with substantial reductions in reported ASRs.
Critical weaknesses
  1. The actual safety fallback of visual masking is unclear. Section 4.1 states that risky instructions trigger a safe fallback, and Appendix E.2 lists several possible strategies (refusal, clarification, confirmation, and read-only execution). However, it is unclear which strategy is actually implemented in the experiments and how the system chooses among them. This matters because these fallback strategies may lead to very different safety and utility outcomes.
  2. Potential limitation of visual masking. Malicious content may overlap with UI elements necessary for the legitimate task. In such cases, masking the entire element could prevent task completion or remove information the agent needs.
  3. Ambiguous results in experiments. Table 5 reports vanilla ASR of 78.4% and BSR of 91.2%, while Table 4 reports undefended completion ASR of 59.6% and Table 6 reports much lower OSWorld benign success rates. I couldn’t determine whether these results use different benchmarks or task subsets. Table 6 also evaluates Gemini-1.5-Pro and Claude-3-Opus, which are not among the 10 main models listed in Section 5.
  4. Lack of judge sensitivity analysis. Both branches rely heavily on LLM-based safety decisions. Appendix specifies Llama-Guard-3-8B for the textual branch and Llama-3.1-8B for visual risk localization, but the paper does not investigate whether the results are sensitive to these choices.
  5. Limited evaluation of coordinated dual-channel attacks. The paper frames as a defense against dual prompt injection. However, experiment separates the evaluation into environmental attacks and user-originated attacks. I would like to see evaluation on an attacker simultaneously manipulates both the textual instruction channel and the visual interface within the same task.

Minor issues:

  1. FPR formulation is wrong. In Section 5, FPR should be FP/(FP + TN).
  2. Overclaim. The paper shows YOLO detector needs fine-tuning, but the framing of ClickShield is training-free. This should be clarified.
  3. Presentation issues. Section 5.2 mis ref Table 8 to Table 6. Appendix D has place holder ?? in Table reference.

Overall, I don’t think this paper is ready for publication.

Questions for authors

Please see above Weakness section.

Reviewer_jW8wOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Unsafe behavior here comes from two sides: harmful instructions and deceptive interface content. ClickShield filters inputs before they reach the base agent. A text judge assesses instruction risk, while a separate perception pipeline detects interface elements, extracts their text, and writes local descriptions. A second judge compares the structured elements against the validated task goal and flags risky regions, which are then masked before the agent acts. The wrapper leaves the agent's policy and the base model untouched.

Critical strengths
  • Filtering interface content before the agent reasons is a clean intervention point, and the structured element representation plus per-region masking decisions make it possible to inspect what was filtered and why.
  • The evaluation goes past the usual comparison against an undefended agent: multiple models and benchmarks, safety and benign outcomes, branch ablations, and latency measurements.
  • Comparing against execution-time monitoring, combined and separately, shows that input filtering adds to later checks rather than duplicating them. The authors also name their failure modes: perception errors, ambiguous interfaces, and useful content lost to masking.
Critical weaknesses
  • The benign results do not establish that masking is precise. Baseline completion is low, so a small absolute drop can still mean losing a large share of the tasks the agent could actually finish. Without uncertainty estimates and per-task changes, the real cost is hard to judge. Overall task success is also not a direct measure of whether risk was localized correctly.
  • Adaptive attacks against the full perception and judging pipeline are barely tested. Separate numbers for user-originated and environment-originated attacks do not establish robustness to an attacker coordinating both channels within one execution.
  • The branch ablations do not isolate detection, text extraction, element descriptions, goal conditioning, and masking, and they leave it unclear how much of the gain comes from refusing instructions rather than from identifying risky visual content.
  • The threat model and trust boundary need to be stated explicitly. The paper treats the user goal as trusted in some places and the user instruction as untrusted in others, and the pseudocode does not say how extra environmental text buffers are handled. Goal-conflict judgments will reject legitimate intermediate steps unless that distinction is spelled out.
  • "Training-free" overstates the case, since some perception components are fine-tuned. What holds is that the base agent needs no retraining; the training cost of the defense itself should be reported separately.
Questions for authors
  • Can you report uncertainty, per-task changes, how many originally successful benign tasks survive, and how often legitimate elements get masked? How much utility remains when attacks are present?
  • Can you evaluate adaptive attacks that coordinate instruction and interface content against the full defense?
  • Can you isolate the effects of the perception components, goal conditioning, and masking, and separate visual filtering from plain instruction refusal?
  • What is trusted in each threat setting, and how are environmental text buffers filtered? How does the system tell a malicious diversion from a legitimate intermediate action?
Reviewer_Cj5hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes ClickShield, a runtime safety wrapper for multimodal computer-use agents. It targets two sources of risk: harmful instructions originating from users and malicious content embedded in the GUI. The method combines a textual risk classifier with a visual pipeline that uses YOLO, OCR, and BLIP-2 to convert screenshots into structured UI elements. A goal-conditioned judge identifies potentially malicious elements, which are then masked before the underlying agent receives the screenshot.

The authors evaluate ClickShield on RiOSWorld, OS-Harm, OSWorld, and CUAHarm. The main reported result is a reduction in adversarial goal completion from 59.6% to 5.4% on RiOSWorld. The paper also includes component ablations, comparisons against prompting-based and runtime defenses, and measurements of benign task performance and inference overhead.

Critical strengths

1. Pre-policy visual sanitization is a useful direction for securing computer-use agents.

The main appeal of ClickShield is its intervention point. Unlike defenses that inspect an agent's reasoning or proposed actions, ClickShield attempts to remove malicious visual content before it influences the base agent's decisions. This is particularly relevant for screenshot-based agents, where deceptive UI elements and embedded instructions are often encountered together. The UI parsing itself builds on existing techniques such as OmniParser (Lu et al., 2024), but using the parsed elements for goal-conditioned safety filtering is a practically useful extension. I also appreciate that the approach can be integrated without modifying the base agent's policy.

2. The empirical results suggest that the two defense branches are complementary.

The evaluation covers multiple agent backbones and both user-originated and environment-originated risks. The RiOSWorld results are particularly encouraging, with reported adversarial goal completion decreasing from 59.6% to 5.4%. The component ablation in Table 8 is also useful: disabling either branch leaves completion ASR above 32%, substantially higher than that of the full system. Although I have concerns about the evaluation discussed below, these results provide initial evidence that combining instruction filtering with visual sanitization is more effective than either component alone. This makes the proposed direction worth investigating further.

Critical weaknesses

1. The claimed security boundary is not sufficiently established, particularly against an adaptive attacker.

The paper argues that decoupling UI perception from the base agent prevents malicious visual content from influencing its reasoning. However, the security of this separation depends heavily on the visual judge and the accuracy of the intermediate element representation. In Appendix E.4, Jvis receives OCR text and element captions originating from the untrusted screen, and a zero-shot LLM determines which elements should be masked. Thus, attacker-controlled content is still passed to another instruction-following model. Using a structured JSON representation does not, by itself, prevent prompt injection against that judge.

This matters because the detector, captioner, and judge form the actual security-critical path. An attacker could attempt to influence the judge through OCR-visible instructions, distribute a payload across multiple elements, or place malicious content in regions that the parser does not recognize. The paper acknowledges some of these limitations but does not evaluate attacks specifically adapted to ClickShield. Moreover, the RiOSWorld results separate user-originated and environmental risks rather than demonstrating robustness to coordinated attacks involving both channels. As a result, the evidence supports effectiveness against the evaluated attacks, but not yet the stronger claim of a reliable trust boundary against dual prompt injection. This distinction is important in light of prior work on visual prompt injection and architectural isolation, including VPI-Bench (Cao et al., 2025) and CaMeLs Can Use Computers Too (Foerster et al., 2026).

2. The experimental comparisons and utility measurements are not sufficiently clear to support the paper's strongest claims.

My main concern is Table 9. The reported ASR values of 53.9% for GuardAgent and 13.0% for MirrorGuard match the UI-TARS-V1.5 results reported in the original MirrorGuard paper (Zhang et al., 2026). In contrast, ClickShield's 5.4% is presented elsewhere as an aggregate over ten different agent backbones on RiOSWorld. The paper does not explain whether these baselines were rerun under the same models, attack instances, and evaluation protocol. If the results were taken from different experimental settings, the comparison does not establish that ClickShield outperforms these defenses.

There are further ambiguities in the safety-utility evaluation. Table 5 reports a BSR of 90.4% for ClickShield, while the OSWorld results in Table 6 show success rates of only around 2–5%, without clearly explaining the different evaluation populations. Section 5 also defines FPR as FP/(FP+TP), which is the false discovery rate rather than the standard false positive rate. Finally, the claim of negligible utility degradation rests on very low baseline OSWorld success rates under a 15-step execution limit. For example, the GPT-4o success rate changes from 5.03% to 4.78%, but without paired task-level outcomes or uncertainty estimates, it is difficult to determine whether masking genuinely preserves normal functionality. These issues materially weaken the evidence for both the comparative advantage and the claimed safety-utility trade-off.

Questions for authors

Clarifications:

  1. Comparison with existing defenses (Table 9). The GuardAgent and MirrorGuard ASR values appear to match the UI-TARS-V1.5 results in Zhang et al. (2026). Were these methods independently evaluated in this work? If so, were the agent backbones, attack instances, step budgets, and success criteria identical to those used for ClickShield? If not, could the authors provide a matched comparison? This is important for assessing the claimed improvement over existing runtime defenses.

  2. Metric definitions and evaluation protocols. Could the authors explain the datasets and evaluation settings underlying Table 5, particularly why its 90.4% BSR differs so substantially from the OSWorld success rates in Table 6? Please also clarify the FPR formula and how the reported FPR values were calculated. For the ASR measurements, is the judge used to assess malicious intention independent of the judges employed by ClickShield, and is goal completion verified through observable environment states?

Additional evidence that could affect my assessment:

  1. Robustness against attacks targeting the firewall. How does ClickShield behave when the attacker knows the defense architecture and explicitly targets the visual judge or UI parser? In particular, I would like to see results for instructions embedded in OCR-detected regions that attempt to manipulate Jvis, as well as attacks placed in regions missed by the detector. Results for coordinated textual and visual attacks would also help substantiate the central dual prompt injection claim. A breakdown of failures caused by missed detection versus incorrect risk judgment would be especially informative.

  2. Benign task preservation. Could the authors provide the exact OSWorld task subset, number of evaluation episodes, and paired task-level results with and without ClickShield? Given the low baseline success rates, it would be useful to report how often the firewall incorrectly masks legitimate UI elements or prevents completion of tasks that the undefended agent would otherwise solve. Such evidence would make the safety-utility claim considerably more convincing.

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

Related papers

Discussion (0)

Sign in to comment.