Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

A Systematization of Security Vulnerabilities in Computer Use Agents

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper claims that ambient web content—a forum post, a PWA install, two config files, a CSV—can drive a sandboxed computer-use agent to execute attacker code.

desk verdict Worth refereeing for the RCE chain and taxonomy, but the CoT case is partly a harness artifact and the environment defaults are under-documented. read the letter →

arxiv 2507.05445 v1 pith:AJN4JMUH submitted 2025-07-07 cs.CR

classification cs.CR
keywords computeruseagentspromptinjectionremotecodeexecutionchain-of-thoughtexposureclickjackinghuman-in-the-loopbypassthreatmodelingagentsecurity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Computer use agents (CUAs) are AI systems that read a screen, plan steps, and act by clicking, typing, and running tools; this paper tries to prove that they share a set of architectural security weaknesses, not just isolated bugs. The core claim is that ambient web content can steer a sandboxed, hardened agent into executing attacker-controlled code, that visual overlays can turn a benign click into a high-privilege action, and that interface framing can coax out hidden chain-of-thought reasoning. The authors demonstrate these with three red-team case studies against a deployed Operator agent and report the same risk classes across other internal CUA deployments. If the claim is right, current safeguards such as sandboxing, human-in-the-loop prompts, and output filters do not establish a reliable security boundary around agent behavior, and secure CUAs will need provenance tracking, deterministic action gating, and containment of reasoning traces.

What carries the argument

The carrying object is the perception–reasoning–action loop every CUA runs: capture a screenshot or DOM snapshot, fuse it with task instructions and memory, generate chain-of-thought steps, emit a tool-level action, then observe the result and repeat. The paper treats this loop as an attack surface, and the central mechanism is a series of trust hops: the model trusts rendered pixels as ground truth, trusts retrieved text on a page as legitimate instruction, trusts file paths and UI labels as honest affordances, and trusts download-tray clicks as user intent. Each hop is individually defensible in normal operation, but adversarial content can exploit them in sequence so that no single action looks malicious. The paper names seven risk classes that arise from these hops—UI deception, remote code execution via action composition, chain-of-thought exposure, human-in-the-loop bypass, indirect prompt injection, identity ambiguity and over-delegation, and content harms—and uses them to organize the case studies.

What would settle it

Run the paper's three attack chains against an unmodified, production-configured CUA in a default sandbox. If a forum post plus PWA install, config-file writes, and a CSV download cannot produce container code execution, or if a desktop file named `admin_only.txt` captures no internal reasoning, then the demonstrated failures do not generalize to real deployments.

Watch

Extended reading notes

Core claim

The central discovery is that CUA compromise is compositional and perceptual: a chain of individually innocuous actions, each plausible on its own, composes into a full sandbox escape. In the paper's end-to-end demonstration, a technical forum post instructs the agent to install a progressive web app, which grants file-system access; the agent then writes a `mimeapps.list` and a `myshell.desktop` file into the host's application directory, downloads a CSV, and clicks it in the download tray, which triggers `sh /home/oai/share/script.sh` inside the container. A separate case shows that a hidden overlay aligned under a visible 'Enter the blog' button makes the agent submit a payment form in an authenticated session, with no semantic check and no audit trail. A third case shows that when a file named `admin_only.txt` appears on the desktop, the agent writes its internal planning traces into it, treating the label as proof that the destination is private. These three failures trace to the same root causes: no input provenance tracking, no binding between what the interface shows and what an action does, and no boundary between the agent's reasoning layer and the writable surfaces around it.

Load-bearing premise

The broad architectural conclusion assumes the red-team environment—particularly the permissive File System Access and PWA behavior used in the RCE chain and the modified `log_cot()` method used to observe CoT leakage—matches how production CUA systems are actually configured.

Editorial extensions

If this is right

  • Sandboxing alone cannot contain a CUA: the agent can be guided to write configuration files, register MIME handlers, and create desktop launchers that turn a file download into code execution inside the container.
  • Human-in-the-loop approval is a probabilistic signal, not a security boundary; reframing, accessibility justifications, and action decomposition can suppress confirmation prompts.
  • Chain-of-thought traces must be treated as privileged execution metadata, with containment and redaction, because interface labels like 'admin only' or 'developer tool' can induce the agent to expose plans that never reach user-facing output.
  • A fixed taxonomy of seven risk classes can serve as the basis for evaluating and red-teaming CUA deployments, since the same classes reproduced across multiple independently built CUA systems.
  • The architectural gaps—missing provenance, weak interface-action binding, and uncontrolled agent memory and delegation—imply that mitigations should live in the orchestration layer (deterministic gating, scoped tokens, ephemeral sessions) rather than in prompt filtering alone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the architectural reading is right, the browser API surface is the highest-leverage control point: restricting File System Access, PWA installation, and MIME/desktop entry writes in agent contexts would break chains like the one demonstrated, likely more reliably than any model-level guardrail.
  • The CoT exposure case used a modified logging routine and a file the agent believed was private; a natural next test is whether an unmodified production agent leaks reasoning when similar developer-tool labels appear in its environment, since the authors did not establish that this occurs without instrumentation.
  • The seven-class taxonomy could be turned into a benchmark by building adversarial environments for each class and measuring bypass rates across vendors and model versions, which would test the paper's assertion that the flaws are architectural rather than vendor-specific.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper systematizes security vulnerabilities in Computer Use Agents (CUAs), identifies seven risk classes, and reports red-team case studies against OpenAI's Operator and additional undisclosed internal CUA deployments. Three core case studies are presented: clickjacking via a visual overlay, an end-to-end remote code execution (RCE) chain driven by indirect prompt injection through a forum post, a PWA install, mimeapps.list/desktop-file creation, and a downloaded CSV that triggers 'Exec=sh /home/oai/share/script.sh' inside the agent container, and chain-of-thought (CoT) leakage via a modified log_cot() helper that writes reasoning to a visible admin_only.txt file. The paper argues that these findings reveal architectural flaws such as missing input provenance, weak interface-action binding, and insufficient control over agent memory and delegation, and it proposes design principles and a CUA-specific security evaluation framework.

Significance. If the findings are reproducible, the RCE chain in Section 5.2 is a significant demonstration that ambient web content can drive a sandboxed agent to execute attacker-controlled code through composed benign-looking actions. The taxonomy of seven risk classes and the treatment of CoT as privileged execution metadata are useful organizing contributions for the emerging CUA security area. The paper also includes a responsible-disclosure notice and proposes concrete mitigation directions. However, the evidence is largely qualitative, the test environments are not fully specified, and the cross-system generalization rests on undisclosed deployments, which currently limits the strength of the paper's central claims.

major comments (4)
  1. [§5.2, §5.4, §4.2.4] The paper does not document the Chromium policy settings or PWA-install permissions in effect during the RCE experiment. The attack chain relies on File System Access and PWA installation, and Section 4.2.4 itself recommends disabling these via DefaultFileSystemReadGuardSetting=2 and DefaultFileSystemWriteGuardSetting=2, implying the test environment may have used permissive defaults. Without stating the baseline configuration and justifying that it matches production Operator deployments, the end-to-end RCE cannot be distinguished from a finding specific to the test harness. Please provide the exact environment configuration (versions, policy settings, PWA permission prompts, HiTL prompts encountered at each step) and argue for its representativeness.
  2. [§5.3] The CoT leakage case study explicitly modifies the agent's internal log_cot() method and redirects CoT output to admin_only.txt. This is an instrumentation of the test harness rather than a demonstration against an unmodified production system. The observed 'belief-induced exposure' may reflect the agent following a developer-provided logging helper, not a vulnerability inherent to stock Operator. The paper should clearly separate observations obtained on the unmodified system from those requiring code modification, and discuss whether the modified helper corresponds to any realistic production feature or deployment scenario.
  3. [§5 (intro), §5.5, §6] The paper's generalization that 'the same classes of vulnerabilities were consistently reproducible' across 'a diverse set of internal CUA deployments' is unsupported because those deployments are undisclosed, with no versions, configurations, or per-system results. This claim is load-bearing for the 'systematization' and 'architectural flaws' conclusions. Please provide a summary table of tested systems (or de-identified descriptions), the specific attacks reproduced on each, and the date/version context, or alternatively temper the generalization to the configurations actually reported.
  4. [§5.2, §4.4] The RCE chain is described as bypassing Human-in-the-Loop (HiTL) safeguards, but the paper never states whether Operator displayed confirmation prompts at any step (e.g., PWA install, file writes, clicking the downloaded CSV) and how those prompts were bypassed. Without this evidence, the 'HiTL bypass' framing is not justified; the actions may simply not have triggered HiTL. Please report whether HiTL was triggered at each step and, if so, how the agent suppressed or circumvented it.
minor comments (6)
  1. [§5.2, §5.4] The string '⁄tildelow/' appears to be a rendering error for '~/' and should be fixed throughout the RCE case study and mitigation text.
  2. [Throughout] The text contains several typos and spacing errors: 'PW A' should be 'PWA', 'ecompassing' in §4.5.3, 'Fianlly' in §4.5.4, 'aslo' in §4.6.3, and 'eve' in §4.5.3.
  3. [§4.6.2] The reference 'Case Study 4.1' should be 'Case Study 5.1', since the clickjacking case study is presented in Section 5.1.
  4. [§5.3] The code snippet for log_cot() is not consistently formatted and uses 'admin only .txt' while the surrounding text says 'admin_only.txt'; please align the snippet with the prose and ensure the path shown matches the experiment.
  5. [Appendix] Several referenced artifacts (e.g., Figures 4, 9, 14, 18, 23, 26) are mentioned but not all are visible in the main text; please either include the figures or clearly state that they are available as supplementary material.
  6. [Table 1] The risk-mapping table shows Case Study 3 as supporting only Risk 3, but the CoT case study also discusses HiTL bypass and output-filtering failures; consider marking those additional risks if supported by the narrative.

Circularity Check

1 steps flagged · score 4.0 of 10

CoT-leakage case is partly self-constructed via the authors' log_cot() modification; RCE and clickjacking remain independent empirical findings.

  1. self definitional [Section 5.3, Case Study: CoT Leakage via Implicit UI Cues (code listing after Section 5.4)]
    "We modified the agent’s internal ‘log cot()‘ method to redirect CoT output into a file named admin_only.txt, visibly placed on the desktop. ... Once this logging hook was active, the model treated the behavior as part of a secure developer workflow and made no effort to conceal the content. This effectively transformed CoT into a covert channel accessible to the user."

    The observed leakage is a deterministic consequence of the test harness, not an independent property of the CUA. The modified log_cot() method unconditionally appends every query and chunk to admin_only.txt via a shell command, so the paper's own instrumentation guarantees that CoT text will appear in that file. The case study then presents this self-constructed channel as evidence for a 'novel failure mode—belief-induced exposure' and uses it to motivate architectural CoT containment recommendations. The demonstration therefore reduces to the authors' own code modification: they created the logging hook that writes CoT, then report that CoT was written.

full rationale

This is an empirical security paper with no fitted parameters, equations, or prediction-vs-fit structure, so most of its claims are not circular. The clickjacking and RCE case studies are external observations against a real CUA environment; even though the RCE chain's representativeness depends on undocumented File System Access and PWA permissions, that is a methodological gap about generalization, not a circular derivation. The only significant circular element is the CoT case, where the authors modify the agent's log_cot() method to write reasoning to admin_only.txt and then treat the resulting file contents as evidence of a vulnerability. One Azure self-citation appears as a platform reference but is not load-bearing, and no self-citation chain forces the central conclusions. The paper's broader architectural claims still rest on independent demonstrations, so the overall circularity is partial and localized rather than total.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central claims rest on the threat model and on the assumption that the tested environment reflects real CUA deployments. The CoT result additionally assumes that externally surfaced reasoning traces are a meaningful security artifact. No numerical free parameters are used.

assumptions (3)
  • domain assumption The threat model limits the adversary to controlling content that the CUA processes, not the platform, model weights, or orchestrator.
    Section 3 defines this boundary. If the attacker had stronger capabilities the taxonomy would be incomplete; if weaker, some findings would not apply.
  • domain assumption The Docker/Operator test configuration is representative of production CUA deployments.
    Sections 5.1, 5.2, and 5.3 generalize from one named public agent plus several undisclosed internal deployments. The paper does not establish that its Chromium policies, PWA permissions, or File System Access settings match default production configurations.
  • domain assumption Externally surfaced CoT artifacts are meaningful and weaponizable even if not verified as the model's true internal reasoning.
    Footnote 2 acknowledges that OpenAI did not confirm the epistemic status of CoT content returned via orchestration APIs, yet the paper's Risk 3 and Case Study 5.3 treat leaked traces as actionable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Systematization of Security Vulnerabilities in Computer Use Agents." pith.science (2026). https://pith.science/paper/AJN4JMUH

@misc{pith2026250705445,
  author       = {Pith},
  title        = {Pith review of: A Systematization of Security Vulnerabilities in Computer Use Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AJN4JMUH}},
  note         = {Machine review of arXiv:2507.05445}
}
read the original abstract

Computer Use Agents (CUAs), autonomous systems that interact with software interfaces via browsers or virtual machines, are rapidly being deployed in consumer and enterprise environments. These agents introduce novel attack surfaces and trust boundaries that are not captured by traditional threat models. Despite their growing capabilities, the security boundaries of CUAs remain poorly understood. In this paper, we conduct a systematic threat analysis and testing of real-world CUAs under adversarial conditions. We identify seven classes of risks unique to the CUA paradigm, and analyze three concrete exploit scenarios in depth: (1) clickjacking via visual overlays that mislead interface-level reasoning, (2) indirect prompt injection that enables Remote Code Execution (RCE) through chained tool use, and (3) CoT exposure attacks that manipulate implicit interface framing to hijack multi-step reasoning. These case studies reveal deeper architectural flaws across current CUA implementations. Namely, a lack of input provenance tracking, weak interface-action binding, and insufficient control over agent memory and delegation. We conclude by proposing a CUA-specific security evaluation framework and design principles for safe deployment in adversarial and high-stakes settings.

Figures

Figures reproduced from arXiv: 2507.05445 by the authors.

Figure 1
Figure 1. Typical CUA system architecture. The seven risks discussed in the paper are annotated next to the specific inputs [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 4
Figure 4. Examples of Human-in-the-Loop (HiTL) bypasses: (a) Sensitive action framed as a benign UI task; (b) On-screen [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 9
Figure 9. Indirect prompt injection via user-generated content and code repositories. The agent interprets adversarially [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figures from the paper (4 more)
Figure 14
Figure 14. Figure 14: End-to-end Remote Code Execution (RCE) via CUA: (a) PWA installation; (b) agent configuration file generation; [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 18
Figure 18. Figure 18: Chain-of-Thought (CoT) leakage during task execution. Internal reasoning appears in intermediate artifacts (a–b), [PITH_FULL_IMAGE:figures/full_fig_p017_18.png]
Figure 23
Figure 23. Figure 23: Clickjacking attack: The CUA is visually misled into triggering an unintended high-privilege action, with no [PITH_FULL_IMAGE:figures/full_fig_p017_23.png]
Figure 26
Figure 26. Figure 26: Form autofill with private data: Despite stating that user confirmation was needed, the CUA extracts and uses [PITH_FULL_IMAGE:figures/full_fig_p018_26.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Mind the Gap: Action Rebinding Attacks against Android GUI Agents

    cs.CR 2026-01 conditional novelty 6.0 of 10

    A zero-permission Android app can redirect a GUI agent's planned tap to a different app by switching the foreground during the agent's reasoning delay.

Reference graph

Works this paper leans on

21 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [1]

    Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments,

    T. Xie, D. Zhang, J. Chen, X. Li, S. Zhao, R. Cao, J. H. Toh, Z. Cheng, D. Shin, F. Lei, Y . Liu, Y . Xu, S. Zhou, S. Savarese, C. Xiong, V . Zhong, and T. Yu, “Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments,” OpenReview, 2024. [Online]. Available: https://openreview.net/forum?id=tN61DTr4Ed

  2. [2]

    Operator system card,

    OpenAI, “Operator system card,” https://openai.com/index/operato r-system-card/, 2025, accessed: 2025-05-10

  3. [3]

    Developing a computer use model,

    Anthropic, “Developing a computer use model,” https://www.anthro pic.com/news/developing-computer-use, 2025, accessed: 2025-05-10

  4. [4]

    Project Astra: Multimodal AI Assistants,

    Google DeepMind, “Project Astra: Multimodal AI Assistants,” https: //deepmind.google/discover/blog/project-astra, 2024, accessed: 2025- 05-21

  5. [5]

    Tell me, what are you most afraid of? Exploring the Effects of Agent Representation on Information Disclosure in Human-Chatbot Interaction

    Y . Liu et al., “Webarena: A realistic web environment for evaluating agents,” arXiv preprint arXiv:2307.12345 , 2023

  6. [6]

    Hello gpt-4o,

    OpenAI, “Hello gpt-4o,” 2024, accessed: 2025-05-10. [Online]. Available: https://openai.com/index/hello-gpt-4o/

  7. [7]

    React: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, and et al., “React: Synergizing reasoning and acting in language models,” arXiv preprint arXiv:2210.03629 , 2022

  8. [8]

    Reflexion: Lan- guage agents with verbal reinforcement learning,

    N. Shinn, N. Scales, S. Gunasekar, and et al., “Reflexion: Lan- guage agents with verbal reinforcement learning,” arXiv preprint arXiv:2303.11366, 2023

Show all 21 references
  1. [9]

    Red teaming language models with language models,

    D. Ganguli et al. , “Red teaming language models with language models,” arXiv preprint arXiv:2202.03286 , 2022

  2. [10]

    Ignore previous prompt: Attack techniques for language models,

    F. Perez and I. Ribeiro, “Ignore previous prompt: Attack techniques for language models,” arXiv preprint arXiv:2211.09527 , 2022

  3. [11]

    Prompt injection attack against llm-integrated applications,

    Y . Liu, G. Deng, Y . Li, K. Wang, Z. Wang, H. Wang, Y . Zheng, Y . Liu, T. Zhang, and Y . Liu, “Prompt injection attack against llm-integrated applications,” arXiv preprint arXiv:2306.05499 , 2023. [Online]. Available: https://arxiv.org/abs/2306.05499

  4. [12]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” arXiv preprint arXiv:2201.11903 , 2022. [Online]. Available: https://arxiv.org/abs/2201.11903

  5. [13]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu et al. , “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,” January 2025, accessed: 2025-05-15. [Online]. Available: https://arxiv.org/abs/2501.12948

  6. [14]

    Webvoyager: Building an end-to-end web agent with large multimodal models,

    X. Ye et al. , “Webvoyager: Building an end-to-end web agent with large multimodal models,” arXiv preprint arXiv:2401.13919 , 2024

  7. [15]

    Llmauditor: A framework for auditing large language models using human-in-the-loop,

    M. Amirizaniani, J. Yao, A. Lavergne, E. S. Okada, A. Chadha, T. Roosta, and C. Shah, “Llmauditor: A framework for auditing large language models using human-in-the-loop,” arXiv preprint arXiv:2402.09346, 2024. [Online]. Available: https://arxiv.org/abs/24 02.09346

  8. [16]

    Preventing ai hallucinations with human-in-the-loop test- ing,

    A. Goala, “Preventing ai hallucinations with human-in-the-loop test- ing,” https://testlio.com/blog/hitl-ai-hallucinations/, 2024, accessed: 2025-05-26

  9. [17]

    Reducing hallucinations in large language models with custom intervention using amazon bedrock agents,

    Amazon Web Services, “Reducing hallucinations in large language models with custom intervention using amazon bedrock agents,” http s://aws.amazon.com/blogs/machine-learning/reducing-hallucination s-in-large-language-models-with-custom-intervention-using-amazo n-bedrock-agents/...

  10. [18]

    Prompt injection attacks in multimodal and tool-augmented language models,

    N. Kassner and et al., “Prompt injection attacks in multimodal and tool-augmented language models,” in Proceedings of the 2024 IEEE Symposium on Security and Privacy (S&P) , 2024

  11. [19]

    ”human in the loop

    Marsh, “”human in the loop” in ai risk management – not a cure-all approach,” https://www.marsh.com/en/services/cyber-risk/insights/h uman-in-the-loop-in-ai-risk-management-not-a-cure-all-approach.ht ml, 2024, accessed: 2025-05-10

  12. [20]

    From object transition to rce in the chrome renderer,

    M. Y . Mo, “From object transition to rce in the chrome renderer,” https://github.blog/security/vulnerability-research/from-object-trans ition-to-rce-in-the-chrome-renderer, Aug. 2024, accessed: 2025-05- 10

  13. [21]

    Announcing the responses api and computer using agent in azure ai foundry,

    Microsoft Azure AI Team, “Announcing the responses api and computer using agent in azure ai foundry,” May 2024, accessed: 2025-05-14. [Online]. Available: https://azure.microsoft.com/en-us/ blog/announcing-the-responses-api-and-computer-using-agent-in-a zure-ai-foundry/ (a) Po...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.