pith. sign in

arxiv: 2605.25435 · v1 · pith:VOEHZLY2new · submitted 2026-05-25 · 💻 cs.AI

Security of OpenClaw Agents: Fundamentals, Attacks, and Countermeasures

Yuntao Wang , Jianle Ba , Han Liu , Yanghe Pan , Jintao Wei , Zhou Su , Tom H. Luan , Linkang Du This is my paper

Pith reviewed 2026-06-29 21:57 UTC · model grok-4.3

classification 💻 cs.AI
keywords OpenClaw agentsLLM agent securitypersistent memory threatsskill poisoningcognitive manipulationcascading failuressupply chain vulnerabilitiesdefense mechanisms
0
0 comments X

The pith

OpenClaw agents' persistent memory and high autonomy enlarge their attack surface to skill poisoning, cognitive manipulation, cascading failures, and supply-chain risks.

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

This survey maps the security landscape for OpenClaw agents, which are continuously running LLM-driven systems equipped with persistent memory, multi-channel interaction, and autonomous skill execution. The authors argue that these features, unlike those in traditional agent systems, combine high-privilege operations with lasting state to create distinct vulnerabilities. They organize the threats into a layered framework that traces risks through reasoning, action execution, and external interactions. Existing defense approaches are reviewed to sketch the current protection options. The work concludes by noting open questions around long-term reliability in such ecosystems.

Core claim

The paper claims that the distinctive architecture of OpenClaw agents—persistent memory paired with autonomous, high-privilege actions—generates a new set of threats including skill poisoning, cognitive manipulation, multi-agent cascading failures, and supply-chain vulnerabilities, which are best understood through a layered categorization spanning reasoning, execution, and interaction phases.

What carries the argument

A layered threat framework that groups vulnerabilities by the stage of agent operation: reasoning, action execution, and external interaction.

If this is right

  • Protections must specifically target persistent memory to prevent skill poisoning from persisting across sessions.
  • Multi-agent deployments require controls to limit failure propagation between connected agents.
  • Supply-chain checks become necessary for any external skills or plugins loaded by the agents.
  • Defenses against cognitive manipulation must operate at the level of the agent's reasoning trace.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same layered structure could be tested against other persistent-memory agent designs to check whether the threats generalize.
  • Quantitative measurement of attack success rates under each layer would turn the framework into a benchmark for new defenses.
  • Integration points with existing operating-system access controls could reduce the privilege surface without altering agent logic.

Load-bearing premise

The collected literature on OpenClaw agent security is complete enough to support a comprehensive layered threat model and defense overview.

What would settle it

Identification of a documented attack on an OpenClaw agent that cannot be placed in any of the three layers of reasoning, execution, or interaction.

Figures

Figures reproduced from arXiv: 2605.25435 by Han Liu, Jianle Ba, Jintao Wei, Linkang Du, Tom H. Luan, Yanghe Pan, Yuntao Wang, Zhou Su.

Figure 1
Figure 1. Figure 1: Organization structure of this survey paper. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: General architecture of OpenClaw agents. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Taxonomy of security threats to OpenClaw agents, which are [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of agent goal hijack threats: (a) indirect prompt injection, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of memory and context poisoning threats: (a) persistent [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of rogue agents: instruction amnesia via context compres [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of tool misuse and exploitation threats: (a) sequential tool [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Rather than directly manipulating the agent’s reasoning [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: Illustration of agentic supply chain vulnerabilities: (a) ClawHub [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Illustration of unexpected code execution: (a) OS command execution [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Illustration of cascading failures: loop amplification & resource [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Illustration of identity and privilege abuse: credential access. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Illustration of insecure inter-agent communication: (a) CIA issues in [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Illustration of human-agent trust exploitation threats: (a) missing [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
read the original abstract

The rapid evolution of large language model (LLM)-driven autonomous agents has given rise to OpenClaw, a new class of open-source agent frameworks that operate as continuously running, skill-augmented systems with persistent memory, multi-channel interaction, and high degrees of autonomy. Such capabilities enable OpenClaw agents to autonomously execute complex, multi-step tasks and interact seamlessly with external applications, but simultaneously introduce a substantially enlarged attack surface. In particular, the combination of high-privilege operations and persistent memory exposes OpenClaw agents to various emerging threats, including skill poisoning, cognitive manipulation, multi-agent cascading failures, and supply-chain vulnerabilities. In this survey, we present a comprehensive study of the security landscape of OpenClaw agents. We first examine the general architecture and key characteristics that distinguish OpenClaw agents from traditional AI agent systems. We categorize existing security and privacy threats into a layered framework and analyze how vulnerabilities arise during agent reasoning, action execution, and external interaction. Representative defense mechanisms are also reviewed to draw the current defense landscape. Finally, several unresolved issues related to the reliability and trustworthiness of OpenClaw ecosystems are discussed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper surveys the security of OpenClaw agents, presented as a new class of open-source LLM-driven autonomous agents distinguished by continuous operation, skill augmentation, persistent memory, multi-channel interaction, and high autonomy. It examines their architecture and characteristics, categorizes threats into a layered framework spanning reasoning, execution, and interaction phases (including skill poisoning, cognitive manipulation, cascading failures, and supply-chain issues), reviews defense mechanisms, and discusses open issues in reliability and trustworthiness.

Significance. If the literature base is shown to be representative of OpenClaw-specific traits, the survey could serve as a useful organizing reference for an emerging area, highlighting how persistent memory and high-privilege operations enlarge the attack surface beyond traditional agents. The low circularity and survey format are strengths, but the value hinges on whether the categorization is grounded in sufficiently direct evidence rather than broad generalization.

major comments (3)
  1. [§3 and §4] §3 (Architecture and Characteristics) and §4 (Layered Threat Framework): the central claim that OpenClaw agents introduce a 'substantially enlarged attack surface' due to persistent memory and high-privilege operations is load-bearing, yet the manuscript does not provide a systematic count or breakdown of reviewed papers that explicitly address these traits versus general LLM-agent literature; this leaves the framework's specificity to OpenClaw under-supported.
  2. [§4] §4 (Threat Categorization): the categorization into reasoning/execution/interaction phases and the four threat types (skill poisoning, cognitive manipulation, multi-agent cascading failures, supply-chain vulnerabilities) generalizes from broader literature; without an explicit discussion of coverage gaps or inclusion criteria for OpenClaw-specific papers, the representativeness assumption identified in the stress-test note remains unaddressed and risks over-generalization.
  3. [§5] §5 (Defense Mechanisms): the review of countermeasures is presented as drawing the 'current defense landscape,' but lacks any quantitative assessment (e.g., number of defenses per threat category or evaluation of their applicability to persistent-memory agents), weakening the claim that the landscape is comprehensively mapped.
minor comments (2)
  1. [Abstract and §2] The abstract and introduction use 'OpenClaw' without an early formal definition or citation to its originating work; adding this in §2 would improve clarity.
  2. [§4] Several threat examples reference external papers but lack consistent citation formatting or DOIs, making it harder to trace the reviewed literature.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on strengthening the grounding of our OpenClaw-specific claims. We respond to each major comment below.

read point-by-point responses

  1. Referee: [§3 and §4] the central claim that OpenClaw agents introduce a 'substantially enlarged attack surface' due to persistent memory and high-privilege operations is load-bearing, yet the manuscript does not provide a systematic count or breakdown of reviewed papers that explicitly address these traits versus general LLM-agent literature; this leaves the framework's specificity to OpenClaw under-supported.

    Authors: We agree that an explicit quantitative breakdown would better substantiate the claim. The reviewed papers were selected for relevance to the distinguishing traits in §3, but no systematic count versus general LLM-agent literature was provided. We will add a table in the revised manuscript breaking down the literature by direct applicability to OpenClaw traits such as persistent memory and high autonomy. revision: yes

  2. Referee: [§4] the categorization into reasoning/execution/interaction phases and the four threat types generalizes from broader literature; without an explicit discussion of coverage gaps or inclusion criteria for OpenClaw-specific papers, the representativeness assumption remains unaddressed and risks over-generalization.

    Authors: We acknowledge that explicit inclusion criteria and gap discussion would improve rigor. The categorization is based on threats applicable to the architecture in §3, but we will revise §4 to add a subsection detailing selection criteria, coverage gaps in OpenClaw-specific papers, and how the framework accounts for the unique traits. revision: yes

  3. Referee: [§5] the review of countermeasures is presented as drawing the 'current defense landscape,' but lacks any quantitative assessment (e.g., number of defenses per threat category or evaluation of their applicability to persistent-memory agents), weakening the claim that the landscape is comprehensively mapped.

    Authors: We agree that quantitative assessment would strengthen the mapping claim. We will revise §5 to include counts of defenses per threat category (via a summary table) and note applicability to persistent-memory and high-privilege agents based on the reviewed works. revision: yes

Circularity Check

0 steps flagged

Survey paper with no derivation chain reducing to self-inputs

full rationale

This is a literature survey that organizes existing threats and defenses for OpenClaw agents into a layered framework by reviewing external papers on LLM agents. No original equations, fitted parameters, predictions, or uniqueness theorems are derived. The architecture description and threat categorization draw directly from cited works without self-definitional loops or load-bearing self-citations that collapse claims back to the paper's own inputs. Representativeness of the literature base is an external-validity issue, not a circularity reduction. The paper is self-contained against external benchmarks as a review.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper is a survey and does not introduce new free parameters, axioms, or invented entities; it reviews existing literature on OpenClaw agent security.

pith-pipeline@v0.9.1-grok · 5751 in / 994 out tokens · 32043 ms · 2026-06-29T21:57:58.794587+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Understanding and mitigating the risks of OpenClaw for non-technical users: A practical guide with Skill

    cs.CR 2026-06 unverdicted novelty 2.0

    This work categorizes seven risks of OpenClaw for non-technical users, provides plain-language mitigations, and supplies a companion Skill to automate security configurations.

Reference graph

Works this paper leans on

72 extracted references · 30 canonical work pages · cited by 1 Pith paper · 9 internal anchors

  1. [1]

    Large model-based agents: State-of-the-art, cooperation paradigms, security and privacy, and future trends,

    Y . Wang, Y . Pan, Z. Su, Y . Deng, Q. Zhao, L. Du, T. H. Luan, J. Kang, and D. Niyato, “Large model-based agents: State-of-the-art, cooperation paradigms, security and privacy, and future trends,”IEEE Communications Surveys & Tutorials, vol. 28, pp. 1906–1949, 2026

    1906

  2. [2]

    Openclaw

    “Openclaw.” https://openclaw.ai/, 2026. Accessed on 2026-03-15

    2026

  3. [3]

    Openclaw PRISM: A zero-fork, defense-in-depth runtime security layer for tool-augmented LLM agents,

    F. Li, “Openclaw PRISM: A zero-fork, defense-in-depth runtime security layer for tool-augmented LLM agents,”arXiv preprint arXiv:2603.11853, pp. 1–23, 2026

    work page arXiv 2026

  4. [4]

    When openclaw AI agents teach each other: Peer learning patterns in the Moltbook community,

    E. Chen, C. Guan, A. Elshafiey, Z. Zhao, J. Zekeri, A. E. Shaibu, and E. O. Prince, “When openclaw AI agents teach each other: Peer learning patterns in the Moltbook community,”arXiv preprint arXiv:2602.14477, pp. 1–7, 2026

    work page arXiv 2026

  5. [5]

    Openclaw: Personal AI assistant

    OpenClaw, “Openclaw: Personal AI assistant.” https://github.com/openc law/openclaw, 2026. Accessed on 2026-03-10

    2026

  6. [6]

    Uncovering security threats and architecting de- fenses in autonomous agents: A case study of openclaw,

    Z. Ying, X. Yang, S. Wu, Y . Song, Y . Qu, H. Li, T. Li, J. Wang, A. Liu, and X. Liu, “Uncovering security threats and architecting de- fenses in autonomous agents: A case study of openclaw,”arXiv preprint arXiv:2603.12644, pp. 1–9, 2026

    work page arXiv 2026

  7. [7]

    A trajectory-based safety audit of clawdbot (openclaw),

    T. Chen, D. Liu, X. Hu, J. Yu, and W. Wang, “A trajectory-based safety audit of clawdbot (openclaw),”arXiv preprint arXiv:2602.14364, pp. 1– 22, 2026

    work page arXiv 2026

  8. [8]

    Agent privilege separation in open- claw: A structural defense against prompt injection,

    D. Cheng and W.-K. Tsao, “Agent privilege separation in open- claw: A structural defense against prompt injection,”arXiv preprint arXiv:2603.13424, pp. 1–6, 2026

    work page arXiv 2026

  9. [9]

    Clawdrain: Exploiting tool-calling chains for stealthy token exhaustion in openclaw agents,

    B. Dong, H. Feng, and Q. Wang, “Clawdrain: Exploiting tool-calling chains for stealthy token exhaustion in openclaw agents,”arXiv preprint arXiv:2603.00902, pp. 1–7, 2026

    work page arXiv 2026

  10. [10]

    From assistant to double agent: Formalizing and benchmarking attacks on openclaw for personalized local AI agent,

    Y . Wang, F. Xu, Z. Lin, G. He, Y . Huang, H. Gao, Z. Niu, S. Lian, and Z. Liu, “From assistant to double agent: Formalizing and benchmarking attacks on openclaw for personalized local AI agent,”arXiv preprint arXiv:2602.08412, pp. 1–11, 2026

    work page arXiv 2026

  11. [11]

    Openclaw vulnerability: Website-to-local agent takeover

    O. Security, “Openclaw vulnerability: Website-to-local agent takeover.” https://www.oasis.security/blog/openclaw-vulnerability, 2026. Accessed on 2026-03-10

    2026

  12. [12]

    Personal AI agents like openclaw are a security night- mare

    Cisco, “Personal AI agents like openclaw are a security night- mare.” https://blogs.cisco.com/ai/personal-ai-agents-like-openclaw-are-a -security-nightmare, 2026. Accessed on 2026-03-10

    2026

  13. [13]

    Clawworm: Self-propagating attacks across LLM agent ecosystems,

    Y . Zhang, Z. Wei, X. Luan, C. Wu, Z. Zhang, J. Wu, H. Wu, H. Chen, J. Sun, and M. Sun, “Clawworm: Self-propagating attacks across LLM agent ecosystems,”arXiv preprint arXiv:2603.15727, pp. 1–18, 2026

    work page arXiv 2026

  14. [14]

    ClawKeeper: Comprehensive safety protection for openclaw agents through skills, plugins, and watchers,

    S. Liu, C. Li, C. Wang, J. Hou, Z. Chen, L. Zhang, Z. Liu, Q. Ye, Y . Hei, X. Zhang,et al., “ClawKeeper: Comprehensive safety protection for openclaw agents through skills, plugins, and watchers,”arXiv preprint arXiv:2603.24414, pp. 1–22, 2026

    work page arXiv 2026

  15. [15]

    AI agents under threat: A survey of key security challenges and future pathways,

    Z. Deng, Y . Guo, C. Han, W. Ma, J. Xiong, S. Wen, and Y . Xiang, “AI agents under threat: A survey of key security challenges and future pathways,”ACM Computing Surveys, vol. 57, no. 7, pp. 1–36, 2025

    2025

  16. [16]

    The emerged security and privacy of LLM agent: A survey with case studies,

    F. He, T. Zhu, D. Ye, B. Liu, W. Zhou, and P. S. Yu, “The emerged security and privacy of LLM agent: A survey with case studies,”ACM Computing Surveys, vol. 58, no. 6, pp. 1–36, 2025

    2025

  17. [17]

    A survey on trustworthy LLM agents: Threats and countermeasures,

    M. Yu, F. Meng, X. Zhou, S. Wang, J. Mao, L. Pan, T. Chen, K. Wang, X. Li, Y . Zhang,et al., “A survey on trustworthy LLM agents: Threats and countermeasures,” inProceedings of ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), pp. 6216–6226, 2025

    2025

  18. [18]

    Security of Internet of agents: Attacks and countermeasures,

    Y . Wang, Y . Pan, S. Guo, and Z. Su, “Security of Internet of agents: Attacks and countermeasures,”IEEE Open Journal of the Computer Society, vol. 6, pp. 1611–1624, 2025

    2025

  19. [19]

    Taming openclaw: Security analysis and mitigation of autonomous LLM agent threats,

    X. Deng, Y . Zhang, J. Wu, J. Bai, S. Yi, Z. Zou, Y . Xiao, R. Qiu, J. Ma, J. Chen,et al., “Taming openclaw: Security analysis and mitigation of autonomous LLM agent threats,”arXiv preprint arXiv:2603.11619, pp. 1–22, 2026

    work page arXiv 2026

  20. [20]

    Automat- ing agent hijacking via structural template injection,

    X. Deng, J. Wu, M. Chen, Y . Xiao, K. Xu, and Q. Li, “Automat- ing agent hijacking via structural template injection,”arXiv preprint arXiv:2602.16958, pp. 1–16, 2026

    work page arXiv 2026

  21. [21]

    Memory poisoning attack and defense on memory based LLM-agents,

    B. Devarangadi Sunil, I. Sinha, P. Maheshwari, S. Todmal, S. Mallik, and S. Mishra, “Memory poisoning attack and defense on memory based LLM-agents,” inProceedings of Advances in Neural Information Processing Systems (NeurIPS), pp. 1–19, 2025

    2025

  22. [22]

    MM-PoisonRAG: Disrupting Multimodal RAG with Local and Global Poisoning Attacks

    H. Ha, Q. Zhan, J. Kim, D. Bralios, S. Sanniboina, N. Peng, K.-W. Chang, D. Kang, and H. Ji, “MM-poisonRAG: Disrupting multimodal RAG with local and global poisoning attacks,”arXiv preprint arXiv:2502.17832, pp. 1–21, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    AI agents need memory control over more context,

    F. Bousetouane, “AI agents need memory control over more context,” arXiv preprint arXiv:2601.11653, pp. 1–32, 2026

    work page arXiv 2026

  24. [24]

    ReCAP: Recursive context-aware reasoning and planning for large language model agents,

    Z. Zhang, T. Chen, W. Xu, A. Pentland, and J. Pei, “ReCAP: Recursive context-aware reasoning and planning for large language model agents,” inAdvances in Neural Information Processing Systems (NeurIPS), pp. 1– 29, 2025

    2025

  25. [25]

    Prompt injection attack to tool selection in LLM agents,

    J. Shi, Z. Yuan, G. Tie, P. Zhou, N. Z. Gong, and L. Sun, “Prompt injection attack to tool selection in LLM agents,” inProceedings of the Network and Distributed System Security Symposium (NDSS), pp. 1–18, 2026

    2026

  26. [26]

    From allies to adversaries: Manipulating LLM tool-calling through adversarial injection,

    R. Zhang, H. Wang, J. Wang, M. Li, Y . Huang, D. Wang, and Q. Wang, “From allies to adversaries: Manipulating LLM tool-calling through adversarial injection,” inProceedings of the Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies, pp. 2009–2028, 2025

    2009

  27. [27]

    Best-of-N jailbreaking,

    J. Hughes, S. Price, A. Lynch, R. Schaeffer, F. Barez, S. Koyejo, H. Sleight, E. Jones, E. Perez, and M. Sharma, “Best-of-N jailbreaking,” inAdvances in Neural Information Processing Systems (NeurIPS), pp. 1– 85, 2025

    2025

  28. [28]

    Great, now write an article about that: The crescendo multi-turn LLM jailbreak attack,

    M. Russinovich, A. Salem, and R. Eldan, “Great, now write an article about that: The crescendo multi-turn LLM jailbreak attack,” inProceed- ings of USENIX Security Symposium (USENIX Security), pp. 1–20, 2025

    2025

  29. [29]

    Exfiltration over web service: T1567

    MITRE ATT&CK, “Exfiltration over web service: T1567.” https://at tack.mitre.org/techniques/T1567/, 2025. Version 1.5, last modified 24 October 2025

    2025

  30. [30]

    Safetydrift: Predicting when AI agents cross the line before they actually do,

    A. Dhodapkar and F. Pishori, “Safetydrift: Predicting when AI agents cross the line before they actually do,”arXiv preprint arXiv:2603.27148, pp. 1–9, 2026

    work page arXiv 2026

  31. [31]

    File and Directory Discovery

    MITRE ATT&CK, “File and Directory Discovery.” https://attack.mitre.o rg/techniques/T1083/, 2025. Technique T1083, Version 1.7, last modified October 24, 2025

    2025

  32. [32]

    LLM06: Sensitive information disclo- sure

    OW ASP GenAI Security Project, “LLM06: Sensitive information disclo- sure.” https://genai.owasp.org/llmrisk2023-24/llm06-sensitive-informati on-disclosure/, 2024. Accessed on 2026-04-28

    2024

  33. [33]

    Unveiling privacy risks in LLM agent memory,

    B. Wang, W. He, S. Zeng, Z. Xiang, Y . Xing, J. Tang, and P. He, “Unveiling privacy risks in LLM agent memory,” inProceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), pp. 25241–25260, 2025

    2025

  34. [34]

    Simple prompt injection attacks can leak personal data observed by LLM agents during task execution,

    M. Alizadeh, Z. Samei, D. Stetsenko, and F. Gilardi, “Simple prompt injection attacks can leak personal data observed by LLM agents during task execution,”arXiv preprint arXiv:2506.01055, pp. 1–25, 2025

    work page arXiv 2025

  35. [35]

    Towards Secure Agent Skills: Architecture, Threat Taxonomy, and Security Analysis

    Z. Li, J. Wu, X. Ling, X. Cui, and T. Luo, “Towards secure agent skills: Architecture, threat taxonomy, and security analysis,”arXiv preprint arXiv:2604.02837, pp. 1–27, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  36. [36]

    Clawhavoc: 341 malicious clawed skills found by the bot they were targeting

    Koi Security, “Clawhavoc: 341 malicious clawed skills found by the bot they were targeting.” https://www.koi.ai/blog/clawhavoc-341-malic ious-clawedbot-skills-found-by-the-bot-they-were-targeting, Feb. 2026. Accessed on 2026-04-28

    2026

  37. [37]

    Snyk finds prompt injection in 36%, 1467 malicious payloads in a toxicskills study of agent skills supply chain compromise

    L. Beurer-Kellner, A. Kudrinskii, M. Milanta, K. B. Nielsen, H. Sarkar, and L. Tal, “Snyk finds prompt injection in 36%, 1467 malicious payloads in a toxicskills study of agent skills supply chain compromise.” 17 https://snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub/, 2026. Accessed on 2026-02-18

    2026

  38. [38]

    BadSkill: Backdoor Attacks on Agent Skills via Model-in-Skill Poisoning

    G. Tie, J. Shi, P. Zhou, and L. Sun, “BadSkill: Backdoor attacks on agent skills via model-in-skill poisoning,”arXiv preprint arXiv:2604.09378, pp. 1–23, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  39. [39]

    Software supply chain

    Singapore Government Standards Portal, “Software supply chain.” https: //info.standards.tech.gov.sg/control-catalog/cybersecurity/sc/, 2026. In- cludes SC-4 Dependency Manifest Version Pinning and SC-6 Depen- dency Installation during Deployment

    2026

  40. [40]

    Defender’s perspective: Dependency confusion and typosquatting attacks

    M. Kydyraliev, “Defender’s perspective: Dependency confusion and typosquatting attacks.” https://slsa.dev/blog/2024/08/dep-confusion-and-t yposquatting, 2024. Accessed on 2026-01-20

    2024

  41. [41]

    Dependency confusion: How I hacked into apple, microsoft and dozens of other companies

    A. Birsan, “Dependency confusion: How I hacked into apple, microsoft and dozens of other companies.” https://medium.com/@alex.birsan/depe ndency-confusion-4a5d60fec610, 2021. Accessed on 2026-01-20

    2021

  42. [42]

    Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale

    Y . Liu, W. Wang, R. Feng, Y . Zhang, G. Xu, G. Deng, Y . Li, and L. Zhang, “Agent skills in the wild: An empirical study of security vulnerabilities at scale,”arXiv preprint arXiv:2601.10338, pp. 1–23, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  43. [43]

    Don’t let the claw grip your hand: A security analysis and defense framework for openclaw,

    Z. Shan, J. Xin, Y . Zhang, and M. Xu, “Don’t let the claw grip your hand: A security analysis and defense framework for openclaw,”arXiv preprint arXiv:2603.10387, pp. 1–12, 2026

    work page arXiv 2026

  44. [44]

    Ingress tool transfer, technique t1105

    MITRE ATT&CK, “Ingress tool transfer, technique t1105.” MITRE ATT&CK, 2025. Version 2.6, last modified October 24, 2025

    2025

  45. [45]

    Obfuscated Files or Information, Technique T1027

    MITRE ATT&CK, “Obfuscated Files or Information, Technique T1027.” https://attack.mitre.org/techniques/T1027/, 2025. Accessed on 2026-04- 28

    2025

  46. [46]

    Malicious OpenClaw skills used to distribute atomic macos stealer

    A. Oliveira, B. Tancio, D. Fiser, P. Lin, and R. Reyes, “Malicious OpenClaw skills used to distribute atomic macos stealer.” Trend Micro Research, 2026. Accessed on 2026-02-18

    2026

  47. [47]

    How OpenClaw AI agent skills are being weaponized

    B. Quintero, “How OpenClaw AI agent skills are being weaponized.” https://blog.virustotal.com/2026/02/from-automation-to-infection-how.h tml, Feb. 2026. VirusTotal Blog

    2026

  48. [48]

    AgentDojo: A dynamic environment to evaluate prompt injection attacks and defenses for LLM agents,

    E. Debenedetti, J. Zhang, M. Balunovic, L. Beurer-Kellner, M. Fischer, and F. Tram`er, “AgentDojo: A dynamic environment to evaluate prompt injection attacks and defenses for LLM agents,” inProceedings of Advances in Neural Information Processing Systems (NeurIPS), vol. 37, pp. 82895–82920, 2024

    2024

  49. [49]

    HAFixAgent: History-aware automated program repair agent,

    Y . Shi, H. Li, B. Adams, and A. E. Hassan, “HAFixAgent: History-aware automated program repair agent,”arXiv preprint arXiv:2511.01047, pp. 1–27, 2025

    work page arXiv 2025

  50. [50]

    LLM02: Insecure Output Handling

    OW ASP Foundation, “LLM02: Insecure Output Handling.” https://gena i.owasp.org/llmrisk/llm02-insecure-output-handling/, 2025. Accessed on 2026-04-28

    2025

  51. [51]

    CWE-78: Improper Neutralization of Special Elements used in an OS Command (’OS Command Injection’)

    MITRE Corporation, “CWE-78: Improper Neutralization of Special Elements used in an OS Command (’OS Command Injection’).” https: //cwe.mitre.org/data/definitions/78.html, 2024. Accessed on 2026-04-28

    2024

  52. [52]

    How code execution drives key risks in agentic AI systems

    J. Irwin and K. Greshake, “How code execution drives key risks in agentic AI systems.” https://developer.nvidia.com/blog/how-code-execu tion-drives-key-risks-in-agentic-ai-systems/, Nov. 2025. Accessed on 2026-04-16

    2025

  53. [53]

    Your Agent, Their Asset: A Real-World Safety Analysis of OpenClaw

    Z. Wang, H. Tu, L. Zhang, H. Chen, J. Wu, X. Liu, Z. Yuan, T. Pang, M. Q. Shieh, F. Liu, Z. Zheng, H. Yao, Y . Zhou, and C. Xie, “Your agent, their asset: A real-world safety analysis of openclaw,”arXiv preprint arXiv:2604.04759, pp. 1–19, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  54. [54]

    Escaping the agent: On ways to bypass openclaw’s security sandbox

    D. Bors, “Escaping the agent: On ways to bypass openclaw’s security sandbox.” https://labs.snyk.io/resources/bypass-openclaw-security-sandb ox/, Feb. 2026. Accessed on 2026-04-14

    2026

  55. [55]

    Cascade: Composing software-hardware attack gadgets for adversarial threat amplification in compound AI systems,

    S. Banerjee, P. Sahu, A. Vahldiek-Oberwagner, J. Sanchez Vicarte, and M. Tiwari, “Cascade: Composing software-hardware attack gadgets for adversarial threat amplification in compound AI systems,”arXiv preprint arXiv:2603.12023, pp. 1–11, 2026

    work page arXiv 2026

  56. [56]

    Zombie agents: Persistent control of self-evolving LLM agents via self-reinforcing injections,

    X. Yang, Y . He, S. Ji, B. Hooi, and J. S. Dong, “Zombie agents: Persistent control of self-evolving LLM agents via self-reinforcing injections,” in ICLR 2026 Workshop on Lifelong Agents: Learning, Aligning, Evolving, pp. 1–14, 2026

    2026

  57. [57]

    Progent: Securing AI Agents with Privilege Control

    T. Shi, J. He, Z. Wang, H. Li, L. Wu, W. Guo, and D. Song, “Pro- gent: Programmable privilege control for LLM agents,”arXiv preprint arXiv:2504.11703, pp. 1–30, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  58. [58]

    Agents of Chaos

    N. Shapira, C. Wendler, A. Yen,et al., “Agents of chaos,”arXiv preprint arXiv:2602.20021, pp. 1–84, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  59. [59]

    LLM10:2025 Unbounded Con- sumption

    OW ASP GenAI Security Project, “LLM10:2025 Unbounded Con- sumption.” https://genai.owasp.org/llmrisk/llm102025-unbounded-consu mption/, 2025. Accessed on 2026-04-18

    2025

  60. [60]

    The role of privacy fatigue in online privacy behavior,

    H. Choi, J. Park, and Y . Jung, “The role of privacy fatigue in online privacy behavior,”Computers in Human Behavior, vol. 81, pp. 42–51, 2018

    2018

  61. [61]

    ClawGuard: A Runtime Security Framework for Tool-Augmented LLM Agents Against Indirect Prompt Injection

    W. Zhao, Z. Li, P. Zhang, and J. Sun, “ClawGuard: A runtime security framework for tool-augmented LLM agents against indirect prompt injection,”arXiv preprint arXiv:2604.11790, pp. 1–19, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  62. [62]

    AgentSys: Secure and dynamic LLM agents through explicit hierarchical memory management,

    R. Wen, H. Li, C. Xiao, and N. Zhang, “AgentSys: Secure and dynamic LLM agents through explicit hierarchical memory management,”arXiv preprint arXiv:2602.07398, pp. 1–21, 2026

    work page arXiv 2026

  63. [63]

    2510.02373 , archivePrefix =

    Q. Wei, T. Yang, Y . Wang, X. Li, L. Li, Z. Yin, Y . Zhan, T. Holz, Z. Lin, and X. Wang, “A-MemGuard: A proactive defense framework for LLM- based agent memory,”arXiv preprint arXiv:2510.02373, pp. 1–27, 2025

    work page arXiv 2025

  64. [64]

    AgentDyn: Are your agent security defenses deployable in real-world dynamic environments?,

    H. Li, R. Wen, S. Shi, N. Zhang, Y . V orobeychik, and C. Xiao, “AgentDyn: Are your agent security defenses deployable in real-world dynamic environments?,” 2026

    2026

  65. [65]

    InjecAgent: Benchmark- ing indirect prompt injections in tool-integrated large language model agents,

    Q. Zhan, Z. Liang, Z. Ying, and D. Kang, “InjecAgent: Benchmark- ing indirect prompt injections in tool-integrated large language model agents,” inFindings of the Association for Computational Linguistics (ACL), pp. 10471–10506, 2024

    2024

  66. [66]

    TeamTNT, Group G0139

    MITRE ATT&CK, “TeamTNT, Group G0139.” MITRE ATT&CK, 2025. Accessed on 2026-05-06

    2025

  67. [67]

    Hacking Auto-GPT and escaping its docker container

    L. Euler, “Hacking Auto-GPT and escaping its docker container.” https: //positive.security/blog/auto-gpt-rce, 2023. Accessed on 2026-04-28

    2023

  68. [68]

    STAC: When innocent tools form dangerous chains to jailbreak LLM agents,

    J.-J. Li, J. He, C. Shang, D. Kulshreshtha, X. Xian, Y . Zhang, H. Su, S. Swamy, and Y . Qi, “STAC: When innocent tools form dangerous chains to jailbreak LLM agents,”arXiv preprint arXiv:2509.25624, pp. 1– 30, 2025

    work page arXiv 2025

  69. [69]

    MalTool: Malicious Tool Attacks on LLM Agents

    Y . Hu, Y . Jia, M. Li, D. Song, and N. Gong, “MalTool: Malicious tool attacks on LLM agents,”arXiv preprint arXiv:2602.12194, pp. 1–34, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  70. [70]

    Parasites in the toolchain: A large-scale analysis of attacks on the MCP ecosystem,

    S. Zhao, Q. Hou, Z. Zhan, Y . Wang, Y . Xie, Y . Guo, L. Chen, S. Li, and Z. Xue, “Parasites in the toolchain: A large-scale analysis of attacks on the MCP ecosystem,” pp. 1–18, 2026

    2026

  71. [71]

    Malicious AI models undermine software supply-chain security,

    A. K. Sood and S. Zeadally, “Malicious AI models undermine software supply-chain security,”Communications of the ACM, vol. 68, no. 6, pp. 62–71, 2025

    2025

  72. [72]

    Detecting malicious entra OAuth Apps with LLM-based permission risk scoring,

    A. Mahara, “Detecting malicious entra OAuth Apps with LLM-based permission risk scoring,”arXiv preprint arXiv:2512.15781, pp. 1–54, 2025

    work page arXiv 2025