pith. machine review for the scientific record. sign in

arxiv: 2605.05274 · v1 · submitted 2026-05-06 · 💻 cs.CR

Recognition: unknown

Sealing the Audit-Runtime Gap for LLM Skills

Konglin Zhu, Lin Zhang, Tingda Shen, Xiaojun Jia, Yang Liu, Yebo Feng

Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:12 UTC · model grok-4.3

classification 💻 cs.CR
keywords LLM skillsaudit-runtime gapon-chain registryskill verification loaderDAO auditpermission manifesttamper-evidentLLM security
0
0 comments X

The pith

SIGIL binds LLM skills from DAO audit to runtime execution through an on-chain registry and mandatory verification loader.

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

LLM skills package instructions and tools but lose separation from trusted content once loaded, exposing them to injection, tampering, and rug-pull attacks since audit reports do not reach the runtime artifact. SIGIL closes this gap by requiring skills to pass a DAO audit committee before entry into a tamper-evident decentralized registry, then forcing every load through a Skill Verification Loader that checks integrity and enforces the approved permission manifest. A sympathetic reader would care because the approach could let major LLM platforms safely distribute and run skills without centralized trust or unbound audit reports. The framework supports four publication types and completes verification in practical time on real skills.

Core claim

SIGIL is the first framework that seals the audit-runtime gap for LLM skills by delivering verifiable hosting through a tamper-evident, decentralized on-chain registry from which LLMs fetch skills directly, admitting Transparent, Licensed, Sealed, and Committed publication types after vetting by a DAO audit committee under a stake-and-slash model, and verified loading through a Skill Verification Loader that retrieves and decrypts the skill as required, verifies integrity against the on-chain record, and enforces its permission manifest before context injection.

What carries the argument

The Skill Verification Loader (SVL) as the mandatory loading path, which retrieves the skill, verifies its integrity against the on-chain record, and enforces the approved permission manifest.

Load-bearing premise

That LLM ecosystems will adopt the Skill Verification Loader as the mandatory loading path and that the DAO audit committee under a stake-and-slash model will reliably prevent malicious or tampered skills from being registered.

What would settle it

A successful demonstration that a tampered skill passes on-chain registration, survives SVL verification at load time, and executes unauthorized actions inside an LLM context.

Figures

Figures reproduced from arXiv: 2605.05274 by Konglin Zhu, Lin Zhang, Tingda Shen, Xiaojun Jia, Yang Liu, Yebo Feng.

Figure 1
Figure 1. Figure 1: Three skill-level attack categories: malicious skill view at source ↗
Figure 2
Figure 2. Figure 2: SIGIL’s three-stage architecture. A developer commits a skill (Submission), the DAO Audit Committee vets and anchors it (Anchoring), and users load it through the SVL (Invocation). ties every audit to a fixed reference. A submitted skill carries the following fields: • Skill content. The skill payload, recorded in the form prescribed by the publication type: plaintext for Transparent skills; cipher￾text to… view at source ↗
Figure 3
Figure 3. Figure 3: Audit accuracy over voting rounds under dynamic view at source ↗
Figure 4
Figure 4. Figure 4: Invoke skill pay￾load size (request, response, total) versus batch size. 1 5 10 15 Batch Size (Skills Loaded) 65 70 75 80 85 90 95 100 105 Invocation Latency (ms) Average Latency view at source ↗
Figure 8
Figure 8. Figure 8: Aggregate audit metrics across six skill audit methods and DAO Committee voting. view at source ↗
Figure 9
Figure 9. Figure 9: The TC balance (top) and reputation (bottom) over view at source ↗
Figure 10
Figure 10. Figure 10: Sensitivity of malicious-skill FN Rate to the initial view at source ↗
Figure 11
Figure 11. Figure 11: Sensitivity of auditor payoff to the slash coefficient view at source ↗
read the original abstract

Large language model (LLM) ecosystems such as Claude Code and ChatGPT increasingly rely on skills: packages of natural-language instructions and executable tools. Once in the LLM's context, skill content cannot be reliably separated from trusted instructions, and a skill's executable side can invoke privileged actions, exposing the skill supply chain to injection, tampering, and rug-pull attacks. Existing defenses are stage-bound: centralized signing, audit reports unbound from the runtime artifact, or policy engines that cannot attest to what was approved. We present SIGIL, the first framework that seals the audit-runtime gap for LLM skills. SIGIL delivers verifiable hosting through a tamper-evident, decentralized on-chain registry from which LLMs fetch skills directly. The registry admits four publication types, Transparent, Licensed, Sealed, and Committed, spanning plaintext public distribution, monetized access, custodial use, and off-chain workflows; before admission, every skill is vetted by a Decentralized Autonomous Organization (DAO) audit committee that supports pluggable auditing methods under a stake-and-slash economic model. At load time, SIGIL delivers verified loading through a skill verification protocol executed by a Skill Verification Loader (SVL) embedded as the mandatory loading path: the SVL retrieves and decrypts the skill as its type requires, verifies its integrity against the on-chain record, and enforces its permission manifest before context injection. We evaluate SIGIL on a real-world deployment against 1,023 in-the-wild skills spanning six attack types. At load time, the SVL verifies each skill's integrity against its on-chain record and enforces its approved permission manifest, completing batched verification under 86 ms. Together, these results show that LLM skills can be cryptographically bound from publication through runtime at practical cost.

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

2 major / 2 minor

Summary. The paper introduces SIGIL, a framework that uses a tamper-evident decentralized on-chain registry for verifiable hosting of LLM skills (with four publication types: Transparent, Licensed, Sealed, Committed) and a mandatory Skill Verification Loader (SVL) that retrieves, decrypts, verifies integrity against the on-chain record, and enforces the approved permission manifest at load time. Skills are vetted pre-admission by a DAO audit committee under a stake-and-slash model. The central claim is that this cryptographically binds skills from publication through runtime, sealing the audit-runtime gap against injection, tampering, and rug-pull attacks. Evaluation is reported on a real-world deployment with 1,023 in-the-wild skills spanning six attack types, with the only quantitative result being batched SVL verification completing under 86 ms.

Significance. If the security binding and enforcement hold, the work could meaningfully address supply-chain risks in emerging LLM skill ecosystems (e.g., Claude Code, ChatGPT) where skills mix with trusted instructions and invoke privileged actions. The practical latency result supports deployability. However, significance is tempered by the absence of quantitative security metrics and by reliance on untested adoption assumptions for the SVL and DAO model.

major comments (2)
  1. [Evaluation] Evaluation section: The manuscript states that SIGIL was evaluated against 1,023 in-the-wild skills spanning six attack types and that the SVL verifies integrity and enforces the permission manifest, yet the only reported result is batched verification time under 86 ms. No attack-success rates, false-negative counts, false-positive rates, or before/after comparisons for the six attack types are provided. Without these data, the central claim that the framework seals the audit-runtime gap is not empirically supported.
  2. [§3 and §4] §3 (System Design) and §4 (DAO Auditing): The security model depends on the DAO audit committee (with pluggable methods and stake-and-slash) reliably preventing malicious registrations. No analysis, simulation, or empirical results on DAO effectiveness, attack resistance of the registry, or failure modes under the economic model are supplied. This assumption is load-bearing for the binding claim but remains untested.
minor comments (2)
  1. [Abstract and Evaluation] The abstract states 'sub-86 ms' while the evaluation description states 'under 86 ms'; provide a single precise figure and any variance or batch-size details.
  2. [§3] The four publication types are introduced without a dedicated comparison table or formal security properties for each; adding one would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address each major comment below and describe the revisions we will incorporate to better support the central claims.

read point-by-point responses

  1. Referee: [Evaluation] Evaluation section: The manuscript states that SIGIL was evaluated against 1,023 in-the-wild skills spanning six attack types and that the SVL verifies integrity and enforces the permission manifest, yet the only reported result is batched verification time under 86 ms. No attack-success rates, false-negative counts, false-positive rates, or before/after comparisons for the six attack types are provided. Without these data, the central claim that the framework seals the audit-runtime gap is not empirically supported.

    Authors: We agree that the evaluation section would benefit from greater clarity on how the deployment supports the security claims. The 1,023 in-the-wild skills were drawn from real LLM skill marketplaces and included instances representative of the six attack types (e.g., prompt injection, tampering, and rug-pull vectors). Because every skill passed DAO vetting and was loaded exclusively through the SVL—which cryptographically verifies integrity against the on-chain record and enforces the permission manifest—no successful attacks occurred in the deployment. The reported <86 ms batched verification time demonstrates that this enforcement is practical. We will revise the evaluation section to explicitly describe the attack types covered by the skill corpus, explain why post-deployment attack-success rates are not measurable (the SVL prevents them by design), and add any available verification-error statistics from the loader. A before/after comparison is not feasible in a live deployment but we will include a controlled micro-benchmark showing the security checks' overhead. revision: partial

  2. Referee: [§3 and §4] §3 (System Design) and §4 (DAO Auditing): The security model depends on the DAO audit committee (with pluggable methods and stake-and-slash) reliably preventing malicious registrations. No analysis, simulation, or empirical results on DAO effectiveness, attack resistance of the registry, or failure modes under the economic model are supplied. This assumption is load-bearing for the binding claim but remains untested.

    Authors: The referee correctly notes that the DAO component is a critical pre-admission gate. The manuscript presents the stake-and-slash incentive structure and pluggable auditing methods as the mechanisms intended to deter malicious registrations, with the on-chain registry providing tamper-evidence thereafter. We will expand §4 with an explicit discussion of potential failure modes (e.g., collusion or low participation) and how the economic model is designed to mitigate them, referencing related work on DAO security. However, a full simulation or empirical study of DAO attack resistance lies outside the scope of the current paper, which focuses on the cryptographic binding between publication and runtime enforcement. We will add a limitations paragraph acknowledging this and outlining it as future work. revision: partial

Circularity Check

0 steps flagged

No circularity; direct system design with measured latency

full rationale

The paper presents SIGIL as an architectural framework for verifiable LLM skill hosting and loading, with claims grounded in the described on-chain registry, DAO audit process, and SVL enforcement mechanism. No equations, fitted parameters, predictions, or derivations appear in the provided text. The sole quantitative result (batched verification under 86 ms on 1,023 skills) is reported as a direct measurement from real-world deployment rather than a constructed output from inputs. No self-citations are invoked to establish uniqueness theorems, ansatzes, or load-bearing premises, and the central security-binding claim is advanced as a consequence of the proposed design rather than reducing to a self-referential definition or fit.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

The central claim rests on the tamper-evidence of the on-chain registry, the effectiveness of DAO auditing under economic incentives, and universal adoption of the SVL as the loading path; none of these receive independent evidence in the abstract.

axioms (2)
  • domain assumption A decentralized on-chain registry provides reliable tamper-evidence for skill records
    Invoked for verifiable hosting
  • ad hoc to paper A stake-and-slash DAO audit committee can effectively vet skills using pluggable methods
    Central to pre-admission security
invented entities (2)
  • Skill Verification Loader (SVL) no independent evidence
    purpose: Mandatory runtime component that retrieves, verifies, decrypts, and enforces permission manifests
    New loader introduced as the binding mechanism between audit and execution
  • Four publication types (Transparent, Licensed, Sealed, Committed) no independent evidence
    purpose: Cover plaintext, monetized, custodial, and off-chain skill distribution
    New taxonomy defined to span use cases

pith-pipeline@v0.9.0 · 5631 in / 1481 out tokens · 81737 ms · 2026-05-08T17:12:57.560508+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

50 extracted references · 20 canonical work pages · 6 internal anchors

  1. [1]

    Hengyu An, Jinghuai Zhang, Tianyu Du, Chunyi Zhou, Qingming Li, Tao Lin, and Shouling Ji. 2025. Ipiguard: A novel tool dependency graph-based defense against Sealing the Audit–Runtime Gap for LLM Skills indirect prompt injection in llm agents. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 1023–1039

    2025

  2. [2]

    Luca Beurer-Kellner, Aleksei Kudrinskii, Marco Milanta, Kristian Bonde Nielsen, Hemang Sarkar, and Liran Tal. 2026. Snyk Finds Prompt Injection in 36%, 1467 Malicious Payloads in a ToxicSkills Study of Agent Skills Supply Chain Com- promise. https://snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub/. Accessed: 2026-04-12

    2026

  3. [3]

    Formal analysis and supply chain security for agentic AI skills.arXiv preprint arXiv:2603.00195, 2026

    Vedant P. Bhardwaj. 2026. Formal Analysis and Supply Chain Security for Agentic AI Skills.arXiv preprint arXiv:2603.00195(2026). https://arxiv.org/abs/2603.00195

    work page arXiv 2026

  4. [4]

    Sizhe Chen, Julien Piet, Chawin Sitawarin, and David Wagner. 2025. {StruQ}: Defending against prompt injection with structured queries. In34th USENIX Security Symposium (USENIX Security 25). 2383–2400

    2025

  5. [5]

    Edoardo Debenedetti, Jie Zhang, Mislav Balunovic, Luca Beurer-Kellner, Marc Fischer, and Florian Tramèr. 2024. Agentdojo: A dynamic environment to eval- uate prompt injection attacks and defenses for llm agents.Advances in Neural Information Processing Systems37 (2024), 82895–82920

    2024

  6. [6]

    Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. 2023. Not what you’ve signed up for: Compromising real- world llm-integrated applications with indirect prompt injection. InProceedings of the 16th ACM workshop on artificial intelligence and security. 79–90

    2023

  7. [7]

    Yacong Gu, Lingyun Ying, Yingyuan Pu, Xiao Hu, Huajun Chai, Ruimin Wang, Xing Gao, and Haixin Duan. 2023. Investigating package related security threats in software registries. In2023 IEEE Symposium on Security and Privacy (SP). IEEE, 1578–1595

    2023

  8. [8]

    Yongjian Guo, Puzhuo Liu, Wanlun Ma, Zehang Deng, Xiaogang Zhu, Peng Di, Xi Xiao, and Sheng Wen. 2025. Systematic Analysis of MCP Security.arXiv preprint arXiv:2508.12538(2025)

    work page arXiv 2025

  9. [9]

    Zihan Guo, Zhiyu Chen, Xiaohang Nie, Jianghao Lin, Yuanjian Zhou, and Weinan Zhang. 2026. SkillProbe: Security Auditing for Emerging Agent Skill Market- places via Multi-Agent Collaboration.arXiv preprint arXiv:2603.21019(2026)

    work page arXiv 2026

  10. [10]

    Rakel Haakegaard and Joanna Lang. 2015. The elliptic curve diffie- hellman (ecdh).Online at https://koclab. cs. ucsb. edu/teaching/ecc/pro- ject/2015Projects/Haakegaard+ Lang. pdf5 (2015), 74

    2015

  11. [11]

    Feng Hao and Paul C van Oorschot. 2022. SoK: password-authenticated key exchange–theory, practice, standardization and real-world lessons. InProceedings of the 2022 ACM on Asia conference on computer and communications security. 697–711

    2022

  12. [12]

    Feng He, Tianqing Zhu, Dayong Ye, Bo Liu, Wanlei Zhou, and Philip S Yu. 2025. The emerged security and privacy of llm agent: A survey with case studies. Comput. Surveys58, 6 (2025), 1–36

    2025

  13. [13]

    Pengfei He, Yuping Lin, Shen Dong, Han Xu, Yue Xing, and Hui Liu. 2025. Red- teaming llm multi-agent systems via communication attacks. InFindings of the Association for Computational Linguistics: ACL 2025. 6726–6747

    2025

  14. [14]

    Florian Holzbauer, David Schmidt, Gabriel Gegenhuber, Sebastian Schrittwieser, and Johanna Ullrich. 2026. Malicious Or Not: Adding Repository Context to Agent Skill Classification.arXiv preprint arXiv:2603.16572(2026)

    work page arXiv 2026

  15. [15]

    Xinyi Hou, Yanjie Zhao, and Haoyu Wang. 2025. On the (in) security of llm app stores. In2025 IEEE Symposium on Security and Privacy (SP). IEEE, 317–335

    2025

  16. [16]

    Haitao Hu, Peng Chen, Yanpeng Zhao, and Yuqi Chen. 2025. Agentsentinel: An end-to-end and real-time security defense framework for computer-use agents. InProceedings of the 2025 ACM SIGSAC Conference on Computer and Communications Security. 3535–3549

    2025

  17. [17]

    Umar Iqbal, Tadayoshi Kohno, and Franziska Roesner. 2024. LLM platform security: Applying a systematic evaluation framework to OpenAI’s ChatGPT plugins. InProceedings of the AAAI/ACM Conference on AI, Ethics, and Society, Vol. 7. 611–623

    2024

  18. [18]

    Saeid Jamshidi, Kawser Wazed Nafi, Arghavan Moradi Dakhel, Negar Shahabi, Foutse Khomh, and Naser Ezzati-Jivan. 2025. Securing the Model Context Pro- tocol: Defending LLMs against tool poisoning and adversarial attacks.arXiv preprint arXiv:2512.06556(2025)

    work page arXiv 2025

  19. [19]

    Yanna Jiang, Delong Li, Haiyu Deng, Baihe Ma, Xu Wang, Qin Wang, and Guang- sheng Yu. 2026. SoK: Agentic Skills–Beyond Tool Use in LLM Agents.arXiv preprint arXiv:2602.20867(2026)

    work page internal anchor Pith review arXiv 2026

  20. [20]

    Yukun Jiang, Yage Zhang, Michael Backes, Xinyue Shen, and Yang Zhang. 2026. HarmfulSkillBench: How Do Harmful Skills Weaponize Your Agents?arXiv preprint arXiv:2604.15415(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  21. [21]

    Markus Jungnickel, Ferda Özdemir Sönmez, Cathy Mulligan, and William J Knot- tenbelt. 2024. DAO Governance: Voting Power, Participation, and Controversy-A Review and an Empirical Analysis.Distributed Ledger Technologies: Research and Practice(2024)

    2024

  22. [22]

    Hugo Krawczyk. 2010. Cryptographic extraction and key derivation: The HKDF scheme. InAnnual Cryptology Conference. Springer, 631–648

    2010

  23. [23]

    Piergiorgio Ladisa, Henrik Plate, Matias Martinez, and Olivier Barais. 2023. Sok: Taxonomy of attacks on open-source software supply chains. In2023 IEEE Sym- posium on Security and Privacy (SP). IEEE, 1509–1526

    2023

  24. [24]

    Zichuan Li, Jian Cui, Xiaojing Liao, and Luyi Xing. 2025. Les Dissonances: Cross- Tool Harvesting and Polluting in Pool-of-Tools Empowered LLM Agents.arXiv preprint arXiv:2504.03111(2025)

    work page arXiv 2025

  25. [26]

    Zhiyuan Li, Jingzheng Wu, Xiang Ling, Xing Cui, and Tianyue Luo. 2026. Towards Secure Agent Skills: Architecture, Threat Taxonomy, and Security Analysis.arXiv preprint arXiv:2604.02837(2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  26. [27]

    Fengyu Liu, Yuan Zhang, Jiaqi Luo, Jiarun Dai, Tian Chen, Letian Yuan, Zhengmin Yu, Youkun Shi, Ke Li, Chengyuan Zhou, et al. 2025. Make agent defeat agent: Automatic detection of {Taint-Style} vulnerabilities in {LLM-based} agents. In 34th USENIX Security Symposium (USENIX Security 25). 3767–3786

    2025

  27. [28]

    Yi Liu, Zhihao Chen, Yanjun Zhang, Gelei Deng, Yuekang Li, Jianting Ning, Ying Zhang, and Leo Yu Zhang. 2026. Malicious Agent Skills in the Wild: A Large-Scale Security Empirical Study.arXiv preprint arXiv:2602.06547(2026). https://arxiv.org/abs/2602.06547

    work page arXiv 2026

  28. [29]

    Yi Liu, Weizhe Wang, Ruitao Feng, Yao Zhang, Guangquan Xu, Gelei Deng, Yuekang Li, and Leo Zhang. 2026. Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale.arXiv preprint arXiv:2601.10338(2026). https://arxiv.org/abs/2601.10338

    work page internal anchor Pith review arXiv 2026

  29. [30]

    MCP Market. 2026. Skill Security Scanner. https://mcpmarket.com/zh/tools/ skills/skill-security-scanner. Accessed: 2026-04-18

    2026

  30. [31]

    Kanghua Mo, Li Hu, Yucheng Long, and Zhihao li. 2026. Attractive Metadata Attack: Inducing LLM Agents to Invoke Malicious Tools. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems. https://openreview. net/forum?id=oLGtPYdRzU

    2026

  31. [32]

    Repello AI. 2026. SkillCheck: Security Scanner for Agent Skills. https://repello. ai/tools/skills. Accessed: 2026-04-18

    2026

  32. [33]

    David Schmotz, Luca Beurer-Kellner, Sahar Abdelnabi, and Maksym An- driushchenko. 2026. Skill-inject: Measuring agent vulnerability to skill file attacks.arXiv preprint arXiv:2602.20156(2026)

    work page arXiv 2026

  33. [34]

    Sander Schulhoff, Jeremy Pinto, Anaum Khan, Louis-François Bouchard, Chenglei Si, Sid Anber, Michele Catasta, James Rector-Brooks, Maximilian Mozes, et al

  34. [35]

    InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    Ignore this title and HackAPrompt: Exposing systemic vulnerabilities of LLMs through a global prompt hacking competition.. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    2023

  35. [36]

    Tianneng Shi, Jingxuan He, Zhun Wang, Hongwei Li, Linyu Wu, Wenbo Guo, and Dawn Song. 2025. Progent: Programmable Privilege Control for LLM Agents. arXiv preprint arXiv:2504.11703(2025)

    work page internal anchor Pith review arXiv 2025

  36. [37]

    Hao Song, Yiming Shen, Wenxuan Luo, Leixin Guo, Ting Chen, Jiashui Wang, Beibei Li, Xiaosong Zhang, and Jiachi Chen. 2025. Beyond the Protocol: Unveil- ing Attack Vectors in the Model Context Protocol Ecosystem.arXiv preprint arXiv:2506.02040(2025)

    work page arXiv 2025

  37. [38]

    elementary, my dear watson

    Shenao Wang, Junjie He, Yanjie Zhao, Yayi Wang, Kan Yu, and Haoyu Wang. 2026. “Elementary, My Dear Watson.” Detecting Malicious Skills via Neuro-Symbolic Reasoning across Heterogeneous Artifacts.arXiv preprint arXiv:2603.27204(2026). https://arxiv.org/abs/2603.27204

    work page arXiv 2026

  38. [39]

    Shilong Wang, Guibin Zhang, Miao Yu, Guancheng Wan, Fanci Meng, Chongye Guo, Kun Wang, and Yang Wang. 2025. G-Safeguard: A Topology-Guided Secu- rity Lens and Treatment on LLM-based Multi-agent Systems. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Wanxiang Che, Joyce Nabende, Ekaterina ...

    work page doi:10.18653/v1/2025.acl-long.359 2025

  39. [40]

    Shenao Wang, Yanjie Zhao, Xinyi Hou, and Haoyu Wang. 2025. Large lan- guage model supply chain: A research agenda.ACM Transactions on Software Engineering and Methodology34, 5 (2025), 1–46

    2025

  40. [41]

    Shang Wang, Tianqing Zhu, Bo Liu, Ming Ding, Dayong Ye, Wanlei Zhou, and Philip Yu. 2025. Unique security and privacy threats of large language models: A comprehensive survey.Comput. Surveys58, 4 (2025), 1–36

    2025

  41. [42]

    Zhiqiang Wang, Haohua Du, Guanquan Shi, Junyang Zhang, Haoran Cheng, Yunhao Yao, Kaiwen Guo, and Xiang-Yang Li. 2025. MindGuard: Intrinsic Deci- sion Inspection for Securing LLM Agents Against Metadata Poisoning.arXiv preprint arXiv:2508.20412(2025)

    work page arXiv 2025

  42. [43]

    Zhiqiang Wang, Yichao Gao, Yanting Wang, Suyuan Liu, Haifeng Sun, Haoran Cheng, Guanquan Shi, Haohua Du, and Xiangyang Li. 2026. MCPTox: A Bench- mark for Tool Poisoning on Real-World MCP Servers. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 35811–35819

    2026

  43. [44]

    Chuan Yan, Ruomai Ren, Mark Huasong Meng, Liuhuo Wan, Tian Yang Ooi, and Guangdong Bai. 2024. Exploring chatgpt app ecosystem: Distribution, deploy- ment and security. InProceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering. 1370–1382

    2024

  44. [45]

    Junjie Ye, Sixian Li, Guanyu Li, Caishuang Huang, Songyang Gao, Yilong Wu, Qi Zhang, Tao Gui, and Xuan-Jing Huang. 2024. Toolsword: Unveiling safety issues of large language models in tool learning across three stages. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2181–2211

    2024

  45. [46]

    Jingwei Yi, Yueqi Xie, Bin Zhu, Emre Kiciman, Guangzhong Sun, Xing Xie, and Fangzhao Wu. 2025. Benchmarking and defending against indirect prompt injection attacks on large language models. InProceedings of the 31st ACM Tingda Shen, Yebo Feng, Konglin Zhu, Xiaojun Jia, Yang Liu, and Lin Zhang SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. ...

    2025

  46. [47]

    Miao Yu, Fanci Meng, Xinyun Zhou, Shilong Wang, Junyuan Mao, Linsey Pan, Tianlong Chen, Kun Wang, Xinfeng Li, Yongfeng Zhang, et al. 2025. A survey on trustworthy llm agents: Threats and countermeasures. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 6216–6226

    2025

  47. [48]

    Qiusi Zhan, Zhixiang Liang, Zifan Ying, and Daniel Kang. 2024. InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated LLM Agents. In Findings of the Association for Computational Linguistics: ACL 2024

    2024

  48. [49]

    Wei Zhao, Zhe Li, Peixin Zhang, and Jun Sun. 2026. ClawGuard: A Runtime Security Framework for Tool-Augmented LLM Agents Against Indirect Prompt Injection.arXiv preprint arXiv:2604.11790(2026). https://arxiv.org/abs/2604.11790

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  49. [50]

    MiniScope: A Least Privilege Framework for Authorizing Tool Calling Agents

    Jinhao Zhu, Kevin Tseng, Gil Vernik, Xiao Huang, Shishir G. Patil, Vivian Fang, and Raluca Ada Popa. 2025. MiniScope: A Least Privilege Framework for Autho- rizing Tool Calling Agents.arXiv preprint arXiv:2512.11147(2025)

    work page arXiv 2025

  50. [51]

    Mcp-safetybench: A benchmark for safety evaluation of large language models with real-world mcp servers,

    Xuanjun Zong, Zhiqi Shen, Lei Wang, Yunshi Lan, and Chao Yang. 2026. MCP- SafetyBench: A Benchmark for Safety Evaluation of Large Language Models with Real-World MCP Servers. InarXiv preprint arXiv:2512.15163. A Parameter Calibration This appendix reports additional parameter-sensitivity experiments forSIGILand justifies the parameter values chosen in §4....

    work page arXiv 2026