acceptodds
Under review as a conference paper at ICLR 2027

EvoRiskBench: An Evolving Benchmark for Runtime Security Risks in Workspace Agents

Abstract

Workspace agents combine large language models with execution harnesses to perform stateful, multi-step tasks that access or modify external resources. Existing benchmarks leave gaps in executable coverage of their runtime security risks, while evolving model capabilities, harnesses, tools, and threats motivate benchmark evolution. We introduce EvoRiskBench, an evolving benchmark organized around the EP–Path–EF framework, which links an initial risk entry point to an one-hop technical effect through an agent-mediated risk path. The framework defines nine entry-point categories and five effect categories; a 20-participant study supports their interpretability and classification consistency on representative cases. Guided by this framework, an automated end-to-end workflow constructs and executes risk cases in isolated environments and independently verifies outcomes using runtime traces and environment states. The benchmark provides a reproducible dataset of 450 adversarial tasks across six scenarios. We evaluate nine Model Harness configurations spanning three models (GPT-5.6 Sol, DeepSeek-V4-Pro-0813, and Claude Opus 5) and three harnesses (Claude Code, Codex, and OpenClaw). Our results reveal substantial vulnerabilities across systems. The most vulnerable configuration, Codex with DeepSeek-V4-Pro-0813, reaches a 68.44% attack success rate (ASR), indicating that configuration of workspace agent is insufficient to ensure secure autonomous execution. ASR varies more across models than harnesses, and harness differences depend on the model. The benchmark cases and evaluation platform are available at https://anonymous.4open.science/r/4B08.

Primary area:
datasets and benchmarks
Keywords:
Workspace Agents, Indirect Prompt Injection, Runtime Security, Risk Taxonomy, Evolving Benchmarks, Execution-Based Evaluation

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_8EnGOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes the EP–Path–EF framework, which connects nine risk entry categories to five technical effect categories through agent-mediated risk paths, together with an automated workflow that iteratively constructs each entry-effect pair into executable cases, yielding a frozen set of 450 tasks across six scenarios. The authors evaluate nine model-harness configurations spanning three models and three harnesses and find that ASR varies more across models than across harnesses, with harness effects depending on the model.

Critical strengths
  1. The risk representation covers the full span from entry point to effect, with nine risk entry categories, agent-mediated risk paths, and five technical effects forming an operational taxonomy. A human expert study supports the interpretability and categorical consistency of these categories on representative scenarios.
  2. Case construction and outcome verification are integrated into an end-to-end automated pipeline, which has practical potential for extending the case set as models and harnesses evolve.
Critical weaknesses
  1. Case construction and test cases are generated entirely by LLMs. Quality control rests on the generator's own rules and replay-based trigger verification, with no human review and no validated LLM judge, so the data quality lacks a credible basis.
  2. The evolving claim in the title lacks empirical support. The authors only state that the workflow can extend the case set and provide no analysis of the evolution process or its results, so whether evolution actually occurs and what it changes cannot be verified.
Questions for authors
  1. What does the case generator take as input, such as seed environments, hand-written examples, or existing benchmarks, and how is diversity ensured within each EP–EF cell?
  2. Claude Opus 5's pass rate is 73.33 on Claude Code and 26.89 and 30.00 on Codex and OpenClaw, while the cross-harness gap for the other two models is only 9 to 14 points. Could the authors explain what causes this gap and whether it also affects the reported ASR?
Reviewer_rFiKOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces EvoRiskBench, a benchmark for evaluating runtime security risks in workspace agents under indirect prompt injection. Its EP → Path → EF framework connects nine risk entry points to five technical effects through agent execution paths. The authors develop an automated workflow for generating, refining, and validating executable cases, producing a frozen evaluation set of 450 tasks across six scenarios. Experiments cover nine combinations of three models and three harnesses, with 4,050 executions in total. The reported aggregate attack success rate is 37.46%, and the task success rate is 60.44%. The results reveal substantial variation across configurations and identify specific risk categories that aggregate scores can obscure.

Critical strengths
  1. The paper addresses an important security problem for agents that operate on files, tools, and external resources. Explicitly including agent communication, subagent delegation, MCP services, and skill invocation provides a useful structure for examining risks across different runtime interfaces.

  2. Combining execution trajectories, system events, and environment states provides a useful foundation for checking whether attacks produce concrete consequences. The decomposition into entry points, execution paths, and effects also supports more informative analysis of where vulnerabilities arise.

  3. Evaluating nine configurations on a shared, frozen case set provides useful evidence about how attack effectiveness varies across systems. Reporting legitimate task completion alongside attack success also recognizes the importance of evaluating both security and utility.

Critical weaknesses
  1. Bypassing safeguards weakens the broader claims about complete agent systems. Appendix A.3 states that unattended runs bypass tool approval prompts and additionally bypass the internal sandbox of Codex CLI, while retaining operating system permissions and platform container isolation. These settings remove safeguards that can prevent harmful actions from taking effect. Consequently, the reported ASRs characterize configurations with reduced protections, and their applicability to agents operating with those protections enabled remains unclear. This limits the paper’s broader claims about the security of complete Model × Harness systems.

  2. The paper insufficiently validates the false-positive behavior of the attack success verifiers. The attack success is verified using verifiers defined for individual cases and supported by runtime evidence. However, the paper does not report an independent calibration against reference labels establishing whether the intended technical effects actually occurred. It is therefore unclear how reliably the verifiers reject unsuccessful attempts, unrelated state changes, or actions that remain within the legitimate task’s authorization boundary. False positives could inflate the reported ASRs and affect comparisons if verification errors vary across configurations or risk categories. An independent audit of sampled success verdicts, complemented by known negative controls, would strengthen the evaluation.

  3. The contribution of the experience repository lacks a controlled ablation. The author introduces an experience repository that summarizes successful cases and failed attempts to guide subsequent case generation. However, the evaluation does not isolate its contribution from execution feedback and iterative refinement. The incremental benefit of accumulating and reusing experience therefore remains unclear. A controlled comparison with the repository removed, while retaining the same generator, target risk pairs, execution feedback, refinement budget, and replay validation criteria, would clarify its contribution.

Questions for authors

See weaknesses

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

The paper creates a benchmark for run-time security risks in a more fine-grained setting and tests a bunch of models for the same.

Critical strengths

(1) the paper is timely.

Critical weaknesses

(1) The writing is really bad. please make it organic and simple to read. The writing should be like an academic paper, not like a report. Additionally, I don't think there is a need for so much math notation in sec. 3.

(2) The novelty is low. AgentCanary already covers most of the fine-grained categories, so it is already clear from earlier papers that there needs to be a fine-grained analysis. Also the categories in this paper may not be exhaustive. I didn't find anything new in the results. This makes me question the novelty.

(3) There are very little concrete details for the benchmark construction. Where did the tasks come from? How much AI was used in the construction? If AI was used, what is the human-AI agreement at each usage? Did the authors validate the eventual tasks? Are the tasks diverse and have enough coverage and do they correspond to real-world tasks?

(4) What are the prescriptions of the authors given the results?

Questions for authors

See above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.