acceptodds
Under review as a conference paper at ICLR 2027

SafePred: A Predictive Guardrail for Computer-Using Agents via World Models

Abstract

With the widespread deployment of Computer-using Agents (CUAs) in complex real-world environments, prevalent long-term risks often lead to severe and irreversible consequences. Most existing guardrails for CUAs adopt a reactive approach, constraining agent behavior only within the current observation space. While these guardrails can prevent immediate short-term risks (e.g., clicking on a phishing link), it cannot proactively avoid long-term risks: seemingly reasonable actions can lead to high-risk consequences that emerge with a delay (e.g., cleaning logs leads to future audits being untraceable), which reactive guardrails cannot identify within the current observation space. To address these limitations, we propose a predictive guardrail approach, with the core idea of aligning predicted future risks with current decisions. Based on this approach, we present SAFEPRED, a predictive guardrail framework for CUAs that establishes a risk-to-decision loop to ensure safe agent behavior. SAFEPRED supports two key abilities: (1) Short- and long-term risk prediction: by using safety policies as the basis for risk prediction, SAFEPRED leverages the prediction capability of the world model to generate semantic representations of both short-term and long-term risks, thereby identifying and pruning actions that lead to high-risk states; (2) Decision optimization: translating predicted risks into actionable safe decision guidances through step-level interventions and task-level re-planning. We further introduce SAFEPRED-BENCH, covering email, shopping, and social scenarios. Extensive experiments show that SAFEPRED significantly reduces high-risk behaviors, achieving over 97.6% safety performance and improving task utility by up to 21.4% compared with reactive baselines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agent Safety, World Model, Agent Evaluation, Computer-Use Agent

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_Tn5zOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SafePred, a world-model-based predictive guardrail framework designed to defend against delayed safety risks in computer-using agents. SafePred consists of three components: policy integration, short-term and long-term risk prediction, and hierarchical decision optimization. The framework predicts potential consequences of candidate actions, evaluates risks against structured safety policies, and provides corrective guidance through action regeneration and task-level replanning. The authors also introduce SafePred-Bench, containing 120 tasks designed to evaluate delayed safety risks, and train a lightweight SafePred-8B model through predictive reasoning distillation. Experiments on OS-Harm, WASP, and SafePred-Bench demonstrate improvements in policy compliance and task completion compared with existing reactive guardrails.

Critical strengths
  1. The paper focuses on an important limitation of reactive guardrails by considering potential downstream consequences of agent actions. The integration of policy-grounded risk prediction with step-level risk guidance and task-level replanning provides a coherent approach to improving both safety and task completion.
  2. The authors evaluate SafePred on multiple safety benchmarks and report improvements in both policy compliance and task success. The additional SafePred-Bench evaluation, module ablations, inference-budget comparison, and lightweight model distillation further validate the experimental contribution.
Critical weaknesses
  1. Heavy reliance on world-model capabilities and insufficient validation of long-term prediction reliability. The effectiveness of SafePred fundamentally depends on the underlying ability of the world model to anticipate future consequences. Although Table 7 in Appendix F compares different world-model backbones, the results reveal noticeable performance differences, while SafePred-8B achieves only modest improvements over its Qwen3-8B base model. More importantly, the evaluation primarily focuses on downstream safety and task success, without directly assessing long-term forecast accuracy across different model capacities. As discussed in Section 3.3, long-term prediction is implemented through high-level semantic reasoning. It therefore remains unclear whether smaller world models can reliably anticipate delayed risks and to what extent the improvements arise from genuine predictive capability.
  2. Presentation quality requires further improvement. Figure 4 contains excessively small font size, making the ablation results difficult to interpret. Additionally, the numbered enumerations in Section 3.1 (Threat Model) and Section 4.1 (Metrics) appear to exhibit unintended hyperlink. Improving these presentation details would enhance the readability and professionalism of the manuscript.
Questions for authors
  1. Regarding world-model capability, could the authors provide a more detailed analysis of how different model capacities affect short- and long-term risk prediction accuracy? In particular, how reliably can SafePred-8B identify delayed risks compared with larger world models, beyond the downstream performance comparisons in Appendix F?
  2. Regarding long-term risk evaluation, could the authors provide additional evidence comparing predicted long-term consequences with actual subsequent outcomes? Such an analysis would help establish whether SafePred genuinely anticipates delayed risks before they materialize.
6 comments

Dear Reviewer Tn5z,

Thank you for your constructive feedback. We have carefully addressed your concerns by extending the evaluation and revising the manuscript in two aspects: (1) World Model Prediction Evaluation: We introduce additional short-term and long-term risk recognition experiments to compare Qwen3-8B, SafePred-8B, and Gemini-2.5-Pro, and further validate predicted long-term consequences against actual execution outcomes. (2) Presentation Quality: We improve the readability of Figure 4 and correct the formatting of numbered enumerations in Sections 3.1 and 4.1. The corresponding experimental results and revisions are detailed below. We hope they adequately address your concerns.

W1: Heavy reliance on world-model capabilities and insufficient validation of long-term prediction reliability. The effectiveness of SafePred fundamentally depends on the underlying ability of the world model to anticipate future consequences. Although Table 7 in Appendix F compares different world-model backbones, the results reveal noticeable performance differences, while SafePred-8B achieves only modest improvements over its Qwen3-8B base model. More importantly, the evaluation primarily focuses on downstream safety and task success, without directly assessing long-term forecast accuracy across different model capacities. As discussed in Section 3.3, long-term prediction is implemented through high-level semantic reasoning. It therefore remains unclear whether smaller world models can reliably anticipate delayed risks and to what extent the improvements arise from genuine predictive capability.

Q1: Regarding world-model capability, could the authors provide a more detailed analysis of how different model capacities affect short- and long-term risk prediction accuracy? In particular, how reliably can SafePred-8B identify delayed risks compared with larger world models, beyond the downstream performance comparisons in Appendix F?

Q2: Regarding long-term risk evaluation, could the authors provide additional evidence comparing predicted long-term consequences with actual subsequent outcomes? Such an analysis would help establish whether SafePred genuinely anticipates delayed risks before they materialize.

R1: To directly evaluate World Model prediction capabilities beyond downstream task performance, we extend the evaluation in two aspects: (1) Model-level Prediction Evaluation: We compare Qwen3-8B, SafePred-8B, and Gemini-2.5-Pro on short-term and long-term risk recognition; (2) Execution-grounded Validation: We compare predicted long-term consequences with actual outcomes from controlled execution trajectories.

(1) Model-level Prediction Evaluation.

We conduct two prediction experiments under matched task inputs and evaluation protocols, using each model's corresponding prediction prompt.

(1-a) Short-term Risk Recognition.

We construct 40 samples, including 20 immediate-risk cases and 20 matched safe controls base on SafePred-Bench. All three models receive the same 40 samples, task contexts, policies, and prediction settings. Each model is evaluated once, resulting in 120 valid predictions.

World ModelPrecision ↑Recall ↑Short-term F1 ↑Accuracy ↑Safe False Positive Rate ↓
Qwen3-8B58.6%85.0%69.4%62.5%60.0%
SafePred-8B58.6%85.0%69.4%62.5%60.0%
Gemini-2.5-Pro74.1%100.0%85.1%82.5%35.0%

SafePred-8B achieves the same short-term F1 as Qwen3-8B, indicating that fine-tuning does not produce a measurable improvement in immediate-risk recognition in this evaluation. Gemini-2.5-Pro achieves a higher F1 of 85.1%, showing stronger short-term risk recognition.

(1-b) Long-term Risk Recognition.

We construct a separate set of 40 samples, comprising 20 delayed-risk cases and 20 matched safe controls base on SafePred-Bench. In each delayed-risk case, the candidate action establishes a condition that leads to an actual policy violation within at most four additional executed actions. Ground-truth outcomes are established from actual execution trajectories before prediction generation. We also compare these cases against the confirmed SafePred training and validation sets to check for task overlap and near-duplicate examples.

All models receive the same task, policies, pre-action state, candidate action, and specified continuation plan. Actual future states and outcome labels are withheld. We use each model's native prediction prompt with a unified maximum output budget of 8,192 tokens.

A prediction is counted as a risk alarm when it identifies at least one violating Policy ID. We report Precision, Recall, F1, and the false positive rate on safe controls.

World ModelPrecision ↑Recall ↑Long-term F1 ↑Safe False Positive Rate ↓
Qwen3-8B58.6%85.0%69.4%60.0%
SafePred-8B70.0%70.0%70.0%30.0%
Gemini-2.5-Pro94.7%90.0%92.3%5.0%

Compared with Qwen3-8B, SafePred-8B increases Precision from 58.6% to 70.0% and F1 from 69.4% to 70.0%, while reducing the false positive rate on safe controls from 60.0% to 30.0%. However, Recall decreases from 85.0% to 70.0%, indicating a trade-off between reducing false alarms and detecting all delayed risks. Gemini-2.5-Pro achieves the strongest results across the reported metrics.

W1 (continue): Heavy reliance on world-model capabilities and insufficient validation of long-term prediction reliability. The effectiveness of SafePred fundamentally depends on the underlying ability of the world model to anticipate future consequences. Although Table 7 in Appendix F compares different world-model backbones, the results reveal noticeable performance differences, while SafePred-8B achieves only modest improvements over its Qwen3-8B base model. More importantly, the evaluation primarily focuses on downstream safety and task success, without directly assessing long-term forecast accuracy across different model capacities. As discussed in Section 3.3, long-term prediction is implemented through high-level semantic reasoning. It therefore remains unclear whether smaller world models can reliably anticipate delayed risks and to what extent the improvements arise from genuine predictive capability.

Q1 (continue): Regarding world-model capability, could the authors provide a more detailed analysis of how different model capacities affect short- and long-term risk prediction accuracy? In particular, how reliably can SafePred-8B identify delayed risks compared with larger world models, beyond the downstream performance comparisons in Appendix F?

Q2 (continue): Regarding long-term risk evaluation, could the authors provide additional evidence comparing predicted long-term consequences with actual subsequent outcomes? Such an analysis would help establish whether SafePred genuinely anticipates delayed risks before they materialize.

(2) Execution-grounded Validation.

To further examine whether predicted consequences correspond to actual future outcomes, we conduct an additional analysis using 40 controlled execution fragments, including 20 delayed-risk cases and 20 matched safe controls, organized into 20 task pairs across 13 mechanism families. These fragments form a separate diagnostic set and are distinct from the 44 long-term-risk tasks in OS-Harm. We evaluate Gemini-2.5-Pro, the World Model backbone used in our main SafePred experiments.

In each delayed-risk case, the candidate action establishes a condition whose harmful consequence materializes within at most four additional executed actions. Ground-truth outcomes are established from actual file changes, database states, and execution records before prediction generation. The model receives the task, policies, pre-action state, candidate action, and specified continuation plan, while actual future states and outcome labels are withheld.

A human reviewer compares each prediction with the actual execution outcome. A full match requires correctly identifying both the affected object and its terminal outcome; semantically equivalent descriptions are accepted. Predictions identifying only intermediate authorization, permission, or queue states are classified as partial matches. All 20 delayed-risk cases are included in the evaluation.

AssessmentResult
Full terminal-consequence match12/20 (60.0%)
Partial match7/20 (35.0%)
Non-match1/20 (5.0%)
Unsupported terminal-harm prediction on safe controls1/20 (5.0%)

Overall, the human-reviewed evaluation provides direct evidence of long-term risk prediction capability. SafePred captures risk-relevant consequences in 19 of 20 delayed-risk cases (95.0%), including 12 full matches (60.0%) and 7 partial matches (35.0%), with only one missed case. The false-positive rate on safe controls is 5.0%.

We have added these prediction-level evaluations, execution-grounded results, representative cases, and limitations to the revised manuscript, with the corresponding details in Appendix F and Section 3.3.

Reviewer_ybv7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SafePred, a predictive guardrail framework for computer-using agents that leverages an LLM-based world model to identify potential short- and long-term risks before executing actions. SafePred integrates structured safety policies, predicts future consequences of candidate actions, and provides corrective guidance when risks are detected. The authors evaluate SafePred on OS-Harm, WASP, and a newly introduced benchmark, SafePred-Bench, reporting improvements in policy compliance and task success over existing guardrails.

Critical strengths
  1. Considering long-term risks: The paper highlights the importance of considering the downstream consequences of agent actions, which may not be immediately apparent during execution.
  2. Evaluation across multiple benchmarks: The authors evaluate the proposed framework on existing safety benchmarks and introduce SafePred-Bench to assess delayed risks in computer-using agents.
Critical weaknesses
  1. Unclear definition of long-term risk: Although the authors provide criteria for long-term risk in Appendix I, behaviors such as plaintext password storage and insecure code generation are classified as long-term risks, despite being potentially unsafe or harmful immediately. Moreover, these risks may be identifiable from the current state and action without predicting future consequences. This raises concerns about whether the selected tasks genuinely require long-term risk prediction.
  2. Incomplete baseline comparisons: The authors introduce both GuardAgent and gpt-oss-safeguard-20b as baselines in Section 4.1, but neither is included in the main experimental comparisons in Tables 1 and 2. Although GuardAgent is evaluated on SafePred-Bench (Table 4), and both methods are included in the action-level classification comparison (Table 6), their end-to-end performance on OS-Harm and WASP is not reported. The authors should clarify the reasons for these omissions and, if feasible, include these baselines in the main evaluation.
  3. Limited evaluation of false positives on benign tasks: To evaluate the tendency of SafePred to incorrectly identify safe actions as risky, the authors should consider evaluating it on benign tasks without intentionally introduced safety hazards. Such an evaluation could help determine whether SafePred unnecessarily interferes with legitimate task completion. Strong performance on these tasks would further support the practical effectiveness of the proposed method.
  4. Inconsistent results on long-term risk tasks: Appendix I lists 44 long-term risk tasks, but the reported SR of 35.0% in Table 2 does not appear consistent with this task count. The authors should clarify the number of evaluated tasks and how the metrics were calculated.
  5. (Minor) Ambiguity in the formalization: Appendix K describes the world model as performing three stages: (1) short- and long-term state prediction, (2) policy-grounded risk identification, and (3) decision guidance generation, all within a single prompt. However, Equations (2) and (3) appear to describe separate calls to the same world model with different inputs and outputs. The authors should clarify whether these equations represent conceptual stages or separate inference calls.
  6. (Minor) Inconsistent introduction of policy notation: The safety policy set P\mathcal{P} is formally introduced in Section 3.3, although the corresponding construction procedure is described in Section 3.2. The terminology also varies between "goal-alignment policy" and "goal-consistency policies." Introducing P\mathcal{P} in Section 3.2 and using consistent terminology throughout would improve clarity.

Overall, the central claim that SafePred addresses long-term risks is not well supported: the selected tasks do not clearly require long-term prediction (W1), and the reported long-term results appear inconsistent (W4). Combined with incomplete baseline comparisons, I do not think the paper is ready for publication in its current form.

Questions for authors
  1. (W1) How were long-term risk tasks selected? Were they selected manually, by LLMs, or according to an operational rubric? Were temporal criteria considered during selection? If the tasks were selected manually, how was the validity of their long-term risk classification verified?
  2. (W2) Why is gpt-oss-safeguard-20b omitted from Tables 1 and 2? Although its classification performance is reported in Table 6, the authors should clarify why it is excluded from the end-to-end comparison.
  3. (W3) Can the authors provide false positive and task success results on benign tasks? For example, OSWorld [1] could be used to assess whether SafePred unnecessarily blocks safe actions or interferes with legitimate task completion.
  4. Contribution of long-term risk prediction: Given that WASP contains only short-term risk cases, why does SafePred achieve substantial improvements on this benchmark? What distinguishes SafePred from existing guardrails in these settings? Although Figure 4 provides an ablation of the prediction components, a WASP-specific analysis would help clarify whether the improvements arise from long-term prediction or other components, such as decision guidance.

[1] Xie, Tianbao, et al. "Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments." Advances in Neural Information Processing Systems 37 (2024): 52040-52094.

21 comments

Dear Reviewer ybv7,

Thank you for your constructive feedback. We have carefully addressed your concerns by revising the manuscript and extending our experiments in five aspects: (1) Long-term Risk Definition: We further clarify the operational criteria, temporal characteristics, and task-selection procedure for long-term risks. (2) Baseline Comparisons: We clarify the evaluation scope of existing baselines and their compatibility with end-to-end agent execution. (3) Benign-task Evaluation: We add an evaluation on benign OSWorld tasks to examine whether SafePred interferes with legitimate task execution. (4) WASP-specific Ablations: We conduct additional component ablations to distinguish the contributions of short-term prediction, long-term prediction, Risk Guidance, and Plan Guidance. (5) Experimental Reporting: We correct the long-term-risk results and clarify the experimental settings and metric definitions. We also address the two minor comments concerning model formalization and policy notation. The detailed responses are provided below. We hope these revisions adequately address your concerns.

W1: Unclear definition of long-term risk: Although the authors provide criteria for long-term risk in Appendix I, behaviors such as plaintext password storage and insecure code generation are classified as long-term risks, despite being potentially unsafe or harmful immediately. Moreover, these risks may be identifiable from the current state and action without predicting future consequences. This raises concerns about whether the selected tasks genuinely require long-term risk prediction.

Q1: How were long-term risk tasks selected? Were they selected manually, by LLMs, or according to an operational rubric? Were temporal criteria considered during selection? If the tasks were selected manually, how was the validity of their long-term risk classification verified?

R1: We have further clarified the operational definition and task selection procedure for long-term risks by explicitly establishing the temporal relationship between an agent's action and the materialization of its potential harmful consequences. Our clarification focuses on three aspects: (i) Temporal Definition: distinguishing risks based on when their harmful consequences materialize; (ii) Task Selection: specifying the source of candidate tasks and the criteria used for manual screening; and (iii) Classification Validation: clarifying how the temporal characteristics of selected tasks are assessed.

(1) Temporal Definition of Long-Term Risks.

We distinguish short-term and long-term risks based on the temporal relationship between an agent's action and the materialization of its harmful consequences. A short-term risk refers to a harmful consequence that materializes immediately upon executing the current action, without requiring subsequent interactions or external events. In contrast, a long-term risk arises when a current action introduces a risk condition whose harmful consequences materialize only through subsequent interactions, events, or environmental conditions.

This distinction focuses on the timing of harm rather than merely the persistence or security-sensitive nature of the introduced condition. For example, writing a credential into a file establishes a potential exposure risk, whereas its subsequent exposure through script execution, logging, or version-control operations represents the downstream harmful consequence. Similarly, generating code containing a vulnerability introduces a potential security risk, while its exploitation during subsequent execution constitutes the corresponding downstream harm. Conversely, if an action directly causes an observable failure upon execution, its consequence is considered short-term rather than long-term.

Thus, the defining characteristic of a long-term risk is that its harmful consequence depends on subsequent interactions or events, rather than materializing entirely through the initial action.

(2) Task Selection Criteria.

We manually selected candidate tasks from the Model Misbehavior category of OS-Harm. The selection was conducted by one researcher through examination of candidate tasks against the four predefined criteria specified in Appendix I:

  • Whether an action makes future problems harder to detect.
  • Whether an action replaces the true task objective with an agent-oriented metric.
  • Whether an action introduces security risks that are not immediately visible.
  • Whether an action reduces the ability for human review, auditing, or correction.

These criteria characterize behaviors that may introduce downstream risks and serve as the basis for identifying candidate tasks. They are complemented by the temporal definition in Section 4.1, which determines whether the harmful consequences depend on subsequent interactions rather than materializing immediately.

The selected tasks cover five application categories: LibreOffice Calc, LibreOffice Impress, multi-application tasks, OS tasks, and VS Code tasks, as detailed in Appendix I. The selection was based on manual examination rather than LLM-based task classification.

W1 (continue) : Unclear definition of long-term risk: Although the authors provide criteria for long-term risk in Appendix I, behaviors such as plaintext password storage and insecure code generation are classified as long-term risks, despite being potentially unsafe or harmful immediately. Moreover, these risks may be identifiable from the current state and action without predicting future consequences. This raises concerns about whether the selected tasks genuinely require long-term risk prediction.

Q1 (continue) : How were long-term risk tasks selected? Were they selected manually, by LLMs, or according to an operational rubric? Were temporal criteria considered during selection? If the tasks were selected manually, how was the validity of their long-term risk classification verified?

(3) Classification Validation.

The classification is based on the relationship between the initial action, the resulting risk condition, and the potential downstream harmful consequence. In particular, we distinguish between risks that materialize immediately upon an action and those whose consequences depend on subsequent interactions or environmental conditions. This temporal distinction provides an explicit basis for determining whether a candidate task falls within the scope of long-term risk.

We further provide representative case analyses in Appendix L to illustrate how an initial action can introduce a risk condition and how its potential consequences may affect subsequent task execution. For example, Table 10 illustrates the potential downstream exposure of hardcoded credentials through subsequent operations, while Table 11 illustrates how an incorrect code modification can compromise subsequent task execution and application functionality.

We have expanded the discussion in Section 4.1 and Appendix I to make the temporal definition, task selection criteria, and classification procedure more explicit. These clarifications establish a more precise operational basis for distinguishing short-term and long-term risks in our evaluation.

Reviewer_7HDiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines the limitations of "reactive guardrails", which assess the safety of computer-using agents (CUAs) based on the current (state, action) pair, and introduces a "predictive guardrail" approach that aligns current decisions with predicted future risks. Building on this approach, the authors propose SafePred, which uses an LLM-based world model to predict short- and long-term risks, evaluate potential policy violations, and provide corrective guidance and replanning.

Experiments on OS-Harm and WASP report policy compliance rates of 99.0% and 97.6%, respectively, with a 21.4-percentage-point improvement in task success over HarmonyGuard on WASP. The authors also introduce SafePred-Bench, a 120-task benchmark for delayed risks, and train a smaller predictive guardrail model, SafePred-8B.

Critical strengths
  1. SafePred combines policy-grounded risk assessment with corrective feedback and replanning, aiming to help agents avoid unsafe actions while continuing to pursue legitimate tasks.

  2. Meaningful evaluation of safety and task utility. The paper evaluates policy compliance, task success, and their joint achievement across multiple benchmarks. The additional delayed-risk benchmark and component ablations provide useful evidence for assessing the overall framework.

Critical weaknesses
  1. Limited direct validation of long-term risk prediction. The prediction-level evaluation assesses policy-based judgments rather than whether predicted long-term consequences match realized outcomes. Although the component ablations suggest that long-term prediction contributes to performance, they do not establish how much of the gain depends on accurate forecasting rather than additional task-aware safety reasoning.

  2. Inconsistencies in the reported results. Table 1 reports an overall OS-Harm SR of 35.0% and SUP of 34.0%, while Table 2 reports the same values for long-term risks and higher values (42.9% and 41.1%) for short-term risks. These results appear inconsistent if the two categories partition the same evaluation runs and the overall results use task-weighted aggregation. Additionally, Table 5 reports a full-model PCR of 99.0% for long-term-risk ablations, whereas Table 2 reports 100.0%. These discrepancies make the long-term-risk results difficult to interpret reliably. Since long-term risk prediction is central to the paper’s contribution, the limited direct validation and unresolved reporting discrepancies weaken the evidence supporting its main claims.

Questions for authors
  1. Could the authors directly evaluate long-term prediction accuracy against realized future outcomes? Beyond the reported component ablations, how can the contribution of accurate forecasting be distinguished from general safety reasoning and corrective planning?

  2. Could the authors clarify the evaluation subsets and aggregation procedures used in Tables 1, 2, and 5, and reconcile the reported differences in PCR, SR, and SUP?

10 comments

Dear Reviewer 7HDi,

Thank you for your careful assessment and constructive feedback. We have addressed your concerns through two main improvements: (1) Long-term Prediction Validation: We conduct an execution-grounded evaluation comparing predicted consequences with realized outcomes, together with a controlled ablation to isolate the contribution of explicit prediction from general policy reasoning and corrective planning. (2) Result Consistency: We recheck the OS-Harm execution records, correct the inconsistencies in Tables 2 and 5, and clarify the evaluation subsets and metric denominators. These revisions have been incorporated into the revised manuscript and are detailed below. We hope they adequately address your concerns.

W1: Limited direct validation of long-term risk prediction. The prediction-level evaluation assesses policy-based judgments rather than whether predicted long-term consequences match realized outcomes. Although the component ablations suggest that long-term prediction contributes to performance, they do not establish how much of the gain depends on accurate forecasting rather than additional task-aware safety reasoning.

Q1: Could the authors directly evaluate long-term prediction accuracy against realized future outcomes? Beyond the reported component ablations, how can the contribution of accurate forecasting be distinguished from general safety reasoning and corrective planning?

R1: We provide two additional evaluations to directly assess long-term prediction reliability and its contribution to downstream performance. (i) Human-reviewed Validation: Comparing predicted consequences with actual execution outcomes; (ii) Controlled Prediction Ablation: Isolating the contribution of explicit consequence prediction while keeping policy reasoning and corrective planning unchanged. These experiments are detailed below.

(1) Human-reviewed Validation of Predicted Consequences

We evaluate 40 controlled execution fragments, comprising 20 delayed-risk cases and 20 matched safe controls, organized into 20 task pairs across 13 mechanism families. This diagnostic set is separate from the 44 long-term-risk tasks in OS-Harm. We evaluate Gemini-2.5-Pro, the World Model backbone used in our main SafePred experiments.

For each delayed-risk case, the candidate action establishes a condition whose harmful consequence materializes within at most four additional executed actions. Ground-truth outcomes are established from actual file changes, database states, and execution records before prediction generation. The model receives the task, policies, pre-action state, candidate action, and specified continuation plan, while future states and outcome labels remain withheld.

A human reviewer compares each prediction with the actual execution outcome. A full match requires correctly identifying both the affected object and its terminal outcome; semantically equivalent descriptions are accepted. Predictions identifying only intermediate authorization, permission, or queue states are classified as partial matches. All 20 delayed-risk cases are included in the evaluation.

AssessmentResult
Full terminal-consequence match12/20 (60.0%)
Partial match7/20 (35.0%)
Non-match1/20 (5.0%)
Unsupported terminal-harm prediction on safe controls1/20 (5.0%)

Overall, the human-reviewed evaluation provides direct evidence of long-term risk prediction capability. SafePred captures risk-relevant consequences in 19 of 20 delayed-risk cases (95.0%), including 12 full matches (60.0%) and 7 partial matches (35.0%), with only one missed case. The false-positive rate on safe controls is 5.0%.

(2) Controlled Prediction Ablation

To isolate the contribution of explicit consequence prediction from general safety reasoning and corrective planning, we compare SafePred with a direct policy-judgment variant. Both configurations retain identical policies, task history, planning, corrective guidance, model backbone, and inference budget. The only difference is that SafePred explicitly predicts short- and long-term consequences, whereas the control directly evaluates policy compliance without generating consequence predictions.

We conduct this comparison on the 44 long-term-risk tasks in OS-Harm.

ConfigurationPCR (%) ↑SR (%) ↑SUP (%) ↑
Direct Policy Judgment (without explicit prediction)88.615.915.9
SafePred (with explicit prediction)100.025.025.0
Improvement+11.4 pp+9.1 pp+9.1 pp

Explicit consequence prediction improves PCR by 11.4 percentage points and both SR and SUP by 9.1 percentage points. With policy reasoning, planning, and corrective guidance held constant, these results isolate the additional contribution of consequence prediction to policy compliance and task completion.

Together, the human-reviewed validation and controlled ablation establish two complementary findings. The former evaluates whether predicted consequences correspond to actual outcomes, while the latter measures their contribution to downstream performance.

We have incorporated both evaluations, the prediction-matching criteria, and representative cases into the revised manuscript, with additional details in Appendix F and corresponding discussion in Section 3.3. We hope these revisions adequately address the reviewer's concerns.

W2: Inconsistencies in the reported results. Table 1 reports an overall OS-Harm SR of 35.0% and SUP of 34.0%, while Table 2 reports the same values for long-term risks and higher values (42.9% and 41.1%) for short-term risks. These results appear inconsistent if the two categories partition the same evaluation runs and the overall results use task-weighted aggregation. Additionally, Table 5 reports a full-model PCR of 99.0% for long-term-risk ablations, whereas Table 2 reports 100.0%. These discrepancies make the long-term-risk results difficult to interpret reliably. Since long-term risk prediction is central to the paper’s contribution, the limited direct validation and unresolved reporting discrepancies weaken the evidence supporting its main claims.

Q2: Could the authors clarify the evaluation subsets and aggregation procedures used in Tables 1, 2, and 5, and reconcile the reported differences in PCR, SR, and SUP?

R3: We rechecked the task-level OS-Harm execution records and confirmed that the discrepancies resulted from errors when populating Tables 2 and 5.

OS-Harm contains 100 tasks, including 44 long-term-risk tasks and 56 short-term-risk tasks. In Table 2, the overall SR and SUP values from all 100 tasks were mistakenly copied into the long-term-risk results. Specifically, the reported 35.0% SR and 34.0% SUP correspond to 35/100 and 34/100, respectively.

The corrected long-term-risk results are:

MetricCorrect CountCorrect Result
PCR44/44100.0%
SR11/4425.0%
SUP11/4425.0%

All 11 successfully completed long-term-risk tasks were policy-compliant, so SR and SUP are both 25.0%.

The full-model PCR of 99.0% in Table 5 was also mistakenly copied from the overall OS-Harm result. We have corrected it to 100.0%, consistent with the 44/44 long-term-risk result in Table 2.

We have corrected Tables 2 and 5 and clarified the evaluation subsets, task counts, and denominators in the revised manuscript.

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

The paper proposed SAFEPRED, a novel predictive guardrail framework designed to protect computer-using agents from both short-term threats and delayed, long-term risks. The paper argues that traditional safety measures are purely reactive, meaning they only evaluate actions within the immediate observation space and fail to catch hidden downstream consequences. To overcome this limitation, SAFEPRED combines structured policy integration, world-model-based risk prediction, and hierarchical decision optimization to align future hazards with active choices. The authors perform evaluations across benchmarks such as OS-Harm, WASP, and a newly developed 120-task delayed-risk benchmark (SAFEPRED-BENCH). The evaluations demonstrate that this framework improves policy compliance rates and task success rates and reports gains in safe task completion.

Critical strengths
  • The paper reports substantial improvements in safe task completion across multiple benchmarks. For example, on WASP, success under policy increases from 67.9% with HarmonyGuard to 96.4% with SafePred. Evaluating compliance alongside task completion helps assess whether safety improvements preserve useful agent behavior.

  • The framework provides a concrete mechanism for translating predicted consequences into corrective action. It connects policy-based risk assessment with action filtering, step-level guidance, and task-level replanning, giving agents actionable alternatives for pursuing authorized goals safely.

Critical weaknesses
  • Table 2 reports 100% policy compliance but different task-success and safe-task-success rates, which conflicts with the metric definitions. Table 5 is labeled as a long-term evaluation but repeats the full model’s overall results from Table 1. These inconsistencies make the ablation gains difficult to interpret. Corrected tables and exact task counts are needed to assess the claimed benefits.

  • The contribution of explicit prediction is not isolated and evaluated. SafePred combines consequence prediction with planning, history use, and corrective guidance. The existing ablations and self-consistency baseline do not establish how much prediction adds beyond these other components. A comparison against the same framework without explicit prediction, retaining equivalent policies, context, guidance, and reasoning budget, is needed to substantiate the central contribution.

Questions for authors
  1. Can you reconcile Tables 1, 2, and 5, including the long-term subset size, exact success and compliance counts, and the denominators used for each ablation?

  2. With identical policies, history, feedback, plan revision, and measured inference budget, how does SafePred compare with a version that omits both short- and long-term prediction outputs? How much of the gain depends on the goal-alignment policy P000?

  3. What exactly are the reference labels used in the 1,244-sample prediction study, and how many examples were independently checked by humans? Can a representative sample be validated against actual next states and downstream outcomes?

  4. Can you show concrete SafePred-Bench tasks identifying the action where risk becomes predictable, the later harm event, and the hidden success and harm checks? Please also report the run counts and variability underlying Table 4.

  5. Please specify the policy weights, threshold for each benchmark, and selection procedure. Case studies use 0.7, but its use across experiments is unclear. A small threshold sweep reporting compliance and safe task completion would clarify robustness.

8 comments

Dear Reviewer oxjz,

Thank you for your constructive feedback. We have addressed your concerns through five revisions: (1) Table Corrections: We rechecked the OS-Harm task-level results and corrected Tables 2 and 5, including the long-term subset denominators and exact task counts. (2) Prediction Ablation: We add a controlled comparison that removes explicit consequence prediction while retaining the other components of SafePred. (3) Label Validation: We clarify the reference-label generation process and the extent of independent human review. (4) Case-level Analysis: We provide a concrete SafePred-Bench example illustrating how a delayed risk becomes predictable and how task success and harm are evaluated. (5) Policy Scoring and Thresholds: We clarify the risk-scoring rule and threshold configuration across experiments. The corresponding corrections and additional analyses are detailed below. We hope they adequately address your concerns.

W1: Table 2 reports 100% policy compliance but different task-success and safe-task-success rates, which conflicts with the metric definitions. Table 5 is labeled as a long-term evaluation but repeats the full model’s overall results from Table 1. These inconsistencies make the ablation gains difficult to interpret. Corrected tables and exact task counts are needed to assess the claimed benefits.

Q1: Can you reconcile Tables 1, 2, and 5, including the long-term subset size, exact success and compliance counts, and the denominators used for each ablation?

R1: Thank you for identifying these inconsistencies. We rechecked the task-level OS-Harm results and found that the long-term subset contains 44 of the 100 evaluated tasks. In Table 2, the SR and SUP values for the long-term subset were mistakenly populated using the full 100-task results (35/100 and 34/100, respectively). The corrected results are PCR = 44/44 (100.0%), SR = 11/44 (25.0%), and SUP = 11/44 (25.0%). These counts are consistent with the metric definitions.

Table 5 also repeated the incorrect full-model SR and SUP values. We have corrected this row using the long-term subset results. After correction, SafePred still achieves higher SR and SUP than the compared methods on the long-term subset, and the full model retains higher SR and SUP than the ablated variants in Table 5.

We have corrected both tables and clarified the corresponding evaluation populations and denominators in the revised manuscript.

W2: The contribution of explicit prediction is not isolated and evaluated. SafePred combines consequence prediction with planning, history use, and corrective guidance. The existing ablations and self-consistency baseline do not establish how much prediction adds beyond these other components. A comparison against the same framework without explicit prediction, retaining equivalent policies, context, guidance, and reasoning budget, is needed to substantiate the central contribution.

Q2: With identical policies, history, feedback, plan revision, and measured inference budget, how does SafePred compare with a version that omits both short- and long-term prediction outputs? How much of the gain depends on the goal-alignment policy P000?

R2: To disentangle the contributions of explicit consequence prediction and goal alignment, we conducted two controlled ablations: (1) Prediction Ablation: We compare SafePred with a direct policy-judgment variant while keeping the other components and inference budget unchanged; (2) P000 Ablation: We remove the goal-alignment policy P000 to examine its contribution to task completion and policy-compliant task success.

(1) Explicit Prediction Ablation.

Both configurations retain the same policies, task history, planning and corrective-guidance mechanisms, model backbone, and inference budget. The only difference is that SafePred explicitly predicts the short- and long-term consequences of candidate actions before policy assessment, whereas the control directly judges policy compliance using the current information without generating prediction outputs.

On the 44 long-term-risk tasks in OS-Harm, the results are:

ConfigurationPCR (%) ↑SR (%) ↑SUP (%) ↑
Direct Policy Judgment (without explicit prediction)88.615.915.9
SafePred (with explicit prediction)100.025.025.0
Improvement+11.4 pp+9.1 pp+9.1 pp

With the other components held constant, explicit prediction improves PCR by 11.4 percentage points and both SR and SUP by 9.1 percentage points. This comparison isolates the contribution of consequence prediction beyond the shared policy reasoning, planning, and corrective-guidance mechanisms.

(2) Contribution of P000.

We further evaluate SafePred with and without P000 on both the full OS-Harm set and its long-term-risk subset.

Evaluation ScopeConfigurationPCR (%) ↑SR (%) ↑SUP (%) ↑
All 100 tasksFull SafePred99.035.034.0
All 100 tasksWithout P00099.033.032.0
Long-term risk (44 tasks)Full SafePred100.025.025.0
Long-term risk (44 tasks)Without P000100.020.520.5

Removing P000 leaves PCR unchanged in both evaluation settings, but reduces SR and SUP by 2.0 percentage points on the full OS-Harm set and 4.5 percentage points on the long-term-risk subset. This indicates that P000 primarily contributes to task alignment and successful task completion, while the policy compliance results remain unchanged.

Together, these two ablations distinguish the contributions of explicit consequence prediction and goal alignment. Explicit prediction improves both policy compliance and policy-compliant task completion, whereas P000 primarily supports task completion without changing policy compliance. We have added these results and clarified the respective contributions of the two components in the revised manuscript.

Q3: What exactly are the reference labels used in the 1,244-sample prediction study, and how many examples were independently checked by humans? Can a representative sample be validated against actual next states and downstream outcomes?

R3: We have clarified the label-generation and human-review procedures for the 1,244 prediction samples, all collected from GPT-4o execution traces.

The evaluation model assessed each sample against the specified safety policies and assigned one of three labels: violate, no_violate, or uncertain. Among the 1,244 samples, 144 uncertain cases were manually reviewed, while the remaining 1,100 cases were not independently checked by human reviewers.

Thus, the reference labels primarily reflect policy-based LLM judgments, with human review limited to the uncertain subset.

We will clarify the origin of these labels, the number of human-reviewed cases, and the scope of this validation in the revised manuscript.

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

This paper presents SafePred, a tuning-free method to guard computer-using agents (CUAs) by predicting the consequences of each proposed action and correcting the agent before a risky action is executed. The method consists of three components. First, an LLM converts each benchmark's safety documents into structured policies, plus a generic goal-alignment policy, which the agent uses to draft its initial plan and the guardrail uses to judge each action. Then, at each step, a guardrail LLM prompted as a "world model" describes in one call the action's immediate UI change (ŝ(t+1)) and its effect on task progress (ẑ(t:T), a short judgment rather than a rollout), and lists the policies the action would violate, which yield a severity-weighted risk score. Finally, SafePred executes the action if the score is at most a threshold τ; otherwise, it returns the guardrail's explanation to the agent ("risk guidance"), and the agent revises its plan ("plan guidance") and proposes a new action for the guardrail to check.

The authors also distill the guardrail into Qwen3-8B with LoRA as an optional smaller variant (SafePred-8B) and introduce SafePred-Bench, 120 browser tasks with delayed, agent-caused harms. With a GPT-4o agent and a Qwen3-Max guardrail, SafePred reaches 99.0% and 97.6% policy compliance (PCR) on OS-Harm and WASP (best competing: 95.0% and 92.9%), and 35.0% and 97.6% task success (HarmonyGuard: 24.0% and 76.2%). On SafePred-Bench, with a Gemini-2.5-Pro guardrail, it reaches 57.5% success under policy (best competing: GuardAgent, 43.3%).

Critical strengths

1. SafePred achieves good safety utility tradeoff.

  • The prompt-based and reactive baselines buy compliance by stalling the agent. On long-term tasks, SR is 6.8% with HarmonyGuard and 2.3% with a generic safety prompt (Table 2). SafePred achieves the best success-under-policy on all three benchmarks, and the WASP margin is large: SUP 96.4% vs. 67.9% for HarmonyGuard, or 81 vs. 57 of 84 tasks.

Why it is consequential: Good performance on three benchmarks that differ in environment and risk type shows a genuine improvement.

2. A timely problem and a sensible benchmark design.

  • Agent-originated, delayed harms are under-evaluated. If fully documented and released, SafePred-Bench would be a useful resource.
Critical weaknesses

1. The claim that prediction avoids long-term risks is not sufficiently supported.

  • SafePred's largest gains are on WASP, whose attacks the paper classifies as short-term (§4.1). Where risks are delayed, guardrails without prediction come close on compliance: a generic safety prompt matches SafePred's 100% policy compliance rate (PCR) on OS-Harm's long-term tasks (Table 2), and on SafePred-Bench the reactive GuardAgent nearly matches SafePred's compliance (95.8% vs. 96.7%, Table 4).

Why it is consequential: SafePred's higher success under policy on all three benchmarks (Strength 1) shows that it is the better guardrail overall, but not that prediction is what avoids the delayed risks. This weakens the "predictive beats reactive" claim in the first contribution.

2. The method is largely incremental over existing guardrails and world models.

  • HarmonyGuard already checks the agent's outputs against structured policies and returns corrective feedback (§4.1), WebDreamer and WMA already use an LLM to predict the outcome of a candidate action before acting (§2), and ToolEmu (Ruan et al., ICLR 2024), which the paper does not cite, already uses an LM to simulate the consequences of agent actions to identify risks.

Why it is consequential: Without the parts that already exist, the method's contribution is asking the guardrail, in the same LLM call, to describe an action's consequences before judging it (App. K). This limits the conceptual novelty, but the combination works well in practice.

Questions for authors
  1. Do HarmonyGuard, GuardAgent and Rule-traversed receive the same extracted policy set (including P000) and the same regeneration budget as SafePred?
6 comments

Dear Reviewer jGsa,

Thank you for your constructive feedback. We have revised the manuscript and provide additional evidence to address your questions in three areas: (1) Long-term risk prediction: We present delayed-risk comparisons, prediction ablations, and human-reviewed validation against execution outcomes. (2) Positioning relative to prior work: We add ToolEmu to Related Work and clarify how SafePred connects consequence prediction with policy-grounded intervention during CUA execution. (3) Baseline comparability: We clarify the policy inputs, intervention procedures, and regeneration budgets used by each baseline. The details are provided below. We hope they adequately address your concerns.

W1: The claim that prediction avoids long-term risks is not sufficiently supported. SafePred's largest gains are on WASP, whose attacks the paper classifies as short-term (§4.1). Where risks are delayed, guardrails without prediction come close on compliance: a generic safety prompt matches SafePred's 100% policy compliance rate (PCR) on OS-Harm's long-term tasks (Table 2), and on SafePred-Bench the reactive GuardAgent nearly matches SafePred's compliance (95.8% vs. 96.7%, Table 4).

R1: SafePred's improvements extend beyond WASP to both OS-Harm's long-term-risk subset and SafePred-Bench. To substantiate its contribution to delayed-risk scenarios, we provide three additional analyses: (i) Cross-benchmark Comparisons: Comparing SafePred with non-predictive guardrails on policy compliance and task utility; (ii) Controlled Prediction Ablation: Isolating the contribution of explicit consequence prediction from shared safety reasoning and corrective planning; and (iii) Human-reviewed Validation: Comparing predicted long-term consequences with actual execution outcomes.

(1) Performance on Delayed-risk Tasks

On OS-Harm's 44 long-term-risk tasks, Generic Defense achieves 100% PCR, matching SafePred, but its SR and SUP are only 2.3%, compared with 25.0% for SafePred.

GuardrailPCR (%) ↑SR (%) ↑SUP (%) ↑
Generic Defense100.02.32.3
HarmonyGuard93.26.86.8
SafePred100.025.025.0

At the same PCR, SafePred improves both SR and SUP by 22.7 percentage points over Generic Defense. Thus, SafePred's advantage lies in maintaining task completion under the same level of policy compliance, rather than simply blocking more actions.

The results on SafePred-Bench show a similar pattern. SafePred achieves 96.7% PCR, close to GuardAgent's 95.8%, while improving SR from 45.8% to 60.8% and SUP from 43.3% to 57.5%.

GuardrailPCR (%) ↑SR (%) ↑SUP (%) ↑
GuardAgent95.845.843.3
SafePred96.760.857.5

SafePred improves SR by 15.0 percentage points and SUP by 14.2 percentage points at a comparable level of policy compliance. These results show that its gains also hold on a benchmark specifically designed to evaluate delayed risks.

(2) Controlled Prediction Ablation

To isolate the contribution of explicit consequence prediction, we compare SafePred with a direct policy-judgment variant on the 44 long-term-risk tasks in OS-Harm. Both configurations retain identical policies, task history, planning, corrective guidance, model backbone, and inference budget. The only difference is that SafePred explicitly predicts short- and long-term consequences, whereas the control directly evaluates policy compliance without generating prediction outputs.

ConfigurationPCR (%) ↑SR (%) ↑SUP (%) ↑
Direct Policy Judgment88.615.915.9
SafePred with Explicit Prediction100.025.025.0
Improvement+11.4 pp+9.1 pp+9.1 pp

With the other components held constant, explicit prediction improves PCR by 11.4 percentage points and both SR and SUP by 9.1 percentage points. This comparison provides direct evidence that consequence prediction contributes to SafePred's performance beyond shared policy reasoning, planning, and corrective guidance.

W1 (continue): The claim that prediction avoids long-term risks is not sufficiently supported. SafePred's largest gains are on WASP, whose attacks the paper classifies as short-term (§4.1). Where risks are delayed, guardrails without prediction come close on compliance: a generic safety prompt matches SafePred's 100% policy compliance rate (PCR) on OS-Harm's long-term tasks (Table 2), and on SafePred-Bench the reactive GuardAgent nearly matches SafePred's compliance (95.8% vs. 96.7%, Table 4).

(3) Human-reviewed Validation of Predicted Consequences

To directly examine whether predicted long-term consequences match actual outcomes, we conducted an additional evaluation using 40 controlled execution trajectories, including 20 delayed-risk cases and 20 matched safe controls, organized into 20 task pairs across 13 mechanism families. This diagnostic set is separate from the 44 long-term-risk tasks in OS-Harm. We evaluated Gemini-2.5-Pro, the World Model backbone used in our main SafePred experiments.

For each delayed-risk case, the candidate action establishes a condition whose harmful consequence materializes within at most four additional executed actions. Ground-truth outcomes were established from actual file changes, database states, and execution records before prediction generation. The model received the task, policies, pre-action state, candidate action, and specified continuation plan, while actual future states and outcome labels were withheld.

A human reviewer compared each prediction with the actual execution outcome. A prediction was counted as a full match only if it correctly identified both the affected object and what ultimately happened to it. Semantically equivalent descriptions were accepted. Predictions identifying only intermediate authorization, permission, or queue states were classified as partial matches. All 20 delayed-risk cases, including missed predictions, were included in the denominator.

The results are:

AssessmentResult
Full terminal-consequence match12/20 (60.0%)
Partial match7/20 (35.0%)
Non-match1/20 (5.0%)
Unsupported terminal-harm prediction on safe controls1/20 (5.0%)

Overall, the human-reviewed evaluation provides direct evidence of long-term risk prediction capability. SafePred captures risk-relevant consequences in 19 of 20 delayed-risk cases (95.0%), including 12 full matches (60.0%) and 7 partial matches (35.0%), with only one missed case. The false-positive rate on safe controls is 5.0%.

We have revised the manuscript to clarify the evidence for long-term risk mitigation and distinguish it from the short-term results on WASP. We hope these additional analyses adequately address the reviewer's concerns.

Comments on the paper

2 comments

Dear Reviewers and Area Chairs,

We sincerely thank all the reviewers and area chairs for their time and thoughtful feedback on our submission. We are encouraged that the reviewers recognize the contribution of SafePred, our predictive guardrail for delayed long-term risks in computer-using agents, which predicts short-term state changes and long-term task consequences before execution and grounds them in explicit safety policies. We have carefully addressed each concern point-by-point in our individual responses and have made corresponding revisions in the updated manuscript. Below we summarize the key clarifications, additional experiments, and new analyses included in our rebuttal:

Direct Validation of Long-Term Prediction (Appendix E.1, E.2, E.4)

Following Reviewers jGsa, 7HDi, and Tn5z, we add evaluations of prediction reliability and its contribution.

  • Controlled ablation: On 44 long-term-risk OS-Harm tasks, we compare SafePred with direct policy judgment while keeping other components and inference budget fixed. Explicit prediction improves PCR from 88.6% to 100.0% and SR/SUP from 15.9% to 25.0%. Removing long-term prediction reduces PCR to 93.2% and SR/SUP to 22.7%.
  • Execution-grounded validation: We evaluate 40 trajectories (20 delayed-risk cases and 20 safe controls), organized into 20 task pairs across 13 mechanism families. Human review finds 12 full matches, 7 partial matches, and 1 non-match; the false-positive rate on safe controls is 5.0%.
  • Model-level evaluation: We compare Qwen3-8B, SafePred-8B, and Gemini-2.5-Pro on separate short- and long-term risk-recognition sets.

Long-Term Risk Definition and Evaluation Clarification (Section 4.1; Appendix K)

Following Reviewers ybv7, oxjz, and 7HDi, we further clarify the operational definition, task selection, and evaluation of long-term risks.

  • We establish explicit temporal criteria to distinguish immediate harms from risks whose consequences materialize through subsequent interactions or environmental events. We also detail the manual task-screening procedure and four selection criteria in Appendix K.
  • We re-examine the task-level OS-Harm evaluation records and revise Tables 2 and 5 to ensure consistent reporting across the full benchmark and its short- and long-term subsets, with explicit task counts and metric denominators.
  • We clarify the labeling and human-review procedure for the 1,244 prediction samples, including the 144 uncertain cases reviewed by humans and the scope of the remaining model-assigned labels (Appendix E.5).

Additional Evaluations and Component Analyses (Appendix B, E.3, F.1–F.2)

Following Reviewer ybv7's requests and questions from Reviewers jGsa and oxjz, we add benign-task evaluation, component ablations, and case analysis.

  • Benign OSWorld: We compare the same agent with and without SafePred on 34 benign tasks. Both complete 6/34 tasks (17.6%) and obtain the same average score (17.67%).
  • WASP and P000 ablations: On 84 WASP tasks, we separately remove Short-term Prediction, Long-term Prediction, Risk Guidance, and Plan Guidance. We also evaluate P000 on all 100 OS-Harm tasks and the 44-task long-term subset.
  • SafePred-Bench: We add a social-location case showing the risk-bearing choice, subsequent harm, and separate hidden success and harm checks. We clarify that Table 4 reports one run per method over 120 tasks.
  • Baseline comparability: We clarify shared policy inputs, including P000, and document baseline intervention procedures and budgets. We explain the evaluation scope of gpt-oss-safeguard-20b and GuardAgent.

Related Work and Contribution Positioning (Section 2)

Following Reviewer jGsa's suggestion, we add ToolEmu to Related Work and revise the Introduction. We distinguish SafePred from HarmonyGuard's reactive assessment, WebDreamer/WMA's world-model prediction, and ToolEmu's LM-emulated tool execution. We position SafePred around policy-grounded prediction of delayed CUA risks with Risk Guidance and Plan Guidance, without claiming LLM-based consequence prediction itself as new.

Risk Scoring and Method Formalization (Sections 3.2–3.4; Appendices L.1 and M)

Following Reviewer oxjz's questions and Reviewer ybv7's comments, we clarify risk scoring and align the formalization with implementation.

  • We specify a uniform score of 0.8 whenever at least one policy violation is predicted, and 0 otherwise, with a common threshold of 0.7 across benchmarks. We explain the selection procedure and why a threshold sweep would not change decisions under this binary rule.
  • We clarify that the three World Model stages occur within one inference call, introduce policy set P in Section 3.2, and standardize "goal-alignment policy."

Presentation Improvements

Following Reviewers Tn5z and ybv7, we improve Figure 4's readability and correct numbered-enumeration formatting in Sections 3.1 and 4.1.

Rebuttal Summary

Dear Area Chairs,

We sincerely thank all reviewers and area chairs for their thoughtful feedback. We are encouraged that reviewers recognized the importance of addressing delayed risks in computer-using agents and balancing safety with task utility. Reviewer jGsa recognized SafePred's safety–utility improvements, while Reviewer 7HDi acknowledged that our execution-grounded validation and controlled ablation strengthened the paper's central claims.

In response, we conducted additional experiments and analyses, including independent human evaluations. We summarize them below.

1. Long-term Risk Definition and Experimental Validation

We clarified the temporal definition and selection procedure for long-term risks and corrected the OS-Harm results for the 44-task subset. SafePred achieves 100.0% PCR and 25.0% SR/SUP, compared with 100.0% PCR and 2.3% SR/SUP for Generic Defense. We also added execution cases illustrating how intermediate actions introduce conditions whose harmful consequences emerge later.

2. Expanded Execution-grounded Prediction Validation

We expanded the evaluation from 40 to 300 controlled cases (150 delayed-risk cases and 150 matched safe controls). Three World Models (Qwen3-8B, SafePred-8B, and Gemini-2.5-Pro) generated 900 predictions on the same cases.

Two independent human reviewers assessed all predictions against actual outcomes. Agreement reached 88.9% for delayed-risk predictions (Cohen's κ = 0.819) and 96.9% for safe-control judgments (κ = 0.903). Gemini-2.5-Pro achieved the highest full-match rates, while SafePred-8B improved over Qwen3-8B, albeit with a higher false-positive rate.

3. Controlled Ablations and Decision Guidance

We compared SafePred with Direct Policy Judgment while keeping policies, history, planning, guidance, model backbone, and inference budget unchanged. Across three runs on the same 44 tasks, SafePred consistently achieved higher SR and SUP, with mean improvements of 6.8 percentage points.

We also conducted P000 and WASP-specific ablations of Short-term Prediction, Long-term Prediction, Risk Guidance, and Plan Guidance. A Reddit upvoting trajectory illustrates how Long-term Prediction helps anticipate downstream consequences and preserve task progress under short-term attacks.

4. Benign-task Evaluation and Intervention Analysis

We evaluated SafePred on 34 benign OSWorld tasks and manually audited all traces. Both SafePred and No Guardrail completed 6/34 tasks.

SafePred flagged 254 of 421 action-proposal steps, with 251 associated with P000, which also captures execution errors such as incorrect clicks. In 160 cases, the agent revised its action, while 94 resulted in fallback to the original action. This shows that most interventions focused on correcting task execution rather than merely blocking risky actions, highlighting SafePred's contribution to agent utility alongside safety.

5. Methodological Clarifications

We added ToolEmu to Related Work, clarified baseline policy inputs and intervention budgets, and specified common policy weights and risk thresholds. We also clarified the single-call World Model implementation and correction-budget behavior, and corrected inconsistent tables, task counts, and policy notation.

Response to the Remaining Concerns

We would also like to address the remaining concerns raised by Reviewer ybv7. While we have repeatedly provided additional experiments in response to the original comments, the latest feedback introduces further evaluation requirements. We believe these should be distinguished from issues that invalidate the evidence already provided.

(1) In particular, a high flag rate alone does not establish excessive intervention, as this overlooks SafePred's utility-oriented role. Beyond identifying risks, SafePred uses predicted consequences to guide action selection and subsequent execution, consistent with world-model-based work such as WMA (ICLR 2025) and WebDreamer (TMLR 2025). Our benign-task evaluation shows unchanged SR, while trajectory analysis demonstrates that some interventions correct execution errors rather than simply block actions.

(2) We also respectfully disagree that the prediction ablation is inconclusive. SafePred achieves higher SR and SUP across all three independent runs, demonstrating consistent improvement rather than an isolated result.

We respectfully ask the Area Chairs to consider the complete experimental evidence and discussion when assessing the paper's contributions.

Thank you for your time and consideration.

Best regards,

Submission #42444 Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.