Cryptography and Security · arXiv:2609.05920 v1 · submitted 5 Sep 2026 · 47 references
Versioned Transitive Dependency-Closure Binding and Operation-Time Effect Governance for Agent Skills: ClosureBound
The paper says
“Agent Skills combine instructions with files, packages, tools, models, and services, so operational identity can exceed a signed directory. Recursive or lazy dependencies may change while root-level evidence remains valid, and different surfaces may reach the same durable effect.”
more of the abstract
We present ClosureBound, a reference monitor that prevents authorization transfer across material changes to this heterogeneous closure. Its resolver commits typed graph nodes and topology. Each grant binds an exact closure root, effect ceiling, purpose/provenance, validity, and epochs. At durability, it re-resolves closure and state, normalizes the operation into an external-effect IR, and admits it only if a joint witness satisfies every bound. Supported equivalent paths share one ceiling. Assuming complete mediation and discovery, authenticated freshness, sound normalization, cryptographic binding, and authoritative linearization, we establish metadata non-authority, closure determinism, version non-inheritance, effect non-amplification, bound-value freshness, and path invariance. We do not establish program equivalence or remote-service honesty. A provider-free implementation matches 40 frozen lifecycle fixtures; 18 kernel contracts and six mutants cover binding and downgrade cases. Full-profile exploration reaches 84,608 states and 530,752 transitions without a declared invariant violation; six weakened profiles yield witnesses. A lexical audit of 549 public Skills (4,872 unique files) finds that 21 of 526 roots with bundled files name every non-manifest path verbatim, 67 contain links resolving outside their roots, and no root declares a frontmatter dependencies field. These observations motivate conservative closure discovery and define concrete targets for broader runtime, interoperability, efficacy, and production validation.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 9 Sep 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. The joint-witness idea is genuinely new and the paper is unusually honest about its limits; the load-bearing gap is that current public Skills do not satisfy the completeness assumption, and the evaluation is self-reported. This line is wrong
3 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 47 references say.
The paper says, at Section 3.3 (Assumption 1) and Section 8.8 (audit)
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
Assumption 1 (complete represented dependency closure) is the load-bearing premise for Theorems 2, 3, and 5. The paper's own static audit (Section 8.8) provides strong evidence that current public Skills do not satisfy it: only 21 of 526 roots with bundled files name every path, 67 contain links resolving outside their root, and no root declares a dependencies field.
more of objection 1
The paper labels this a 'deployment obligation' (Section 11.2) and acknowledges the gap in Section 8.8. This is not an internal inconsistency, but the paper does not discuss any concrete method for a deployment to achieve or conservatively approximate completeness for arbitrary Skills (e.g., through static analysis, runtime instrumentation, or conservative containers). Without such guidance, the practical reach of the theorems is severely limited. I suggest adding a paragraph in Section 11.2 or Section 12 that outlines at least one plausible practical strategy (e.g., pinning all recursive dependencies at activation and rejecting any dynamic load that cannot be fastened to a known node) and identifying what future work would be needed to validate that strategy across diverse Skill types.
The paper says, at Section 8.6 (bounded exploration) and Section 8.8 (audit)
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The bounded exploration (84,608 states, 530,752 transitions) is exhaustive for the encoded finite abstraction but abstracts away package-manager behavior, remote-service changes, and dynamic code loading. Combined with the static audit's finding that real Skills routinely have unrepresented dependencies, the exploration does not demonstrate that the monitor can scale to realistic incomplete closures.
more of objection 2
The paper should more clearly state the gap between the closed-world model used in the exploration and the open-world nature of Skill execution, and discuss whether any of the mutants or profiles capture the omission of a dependency (as opposed to a changed version).
The paper says, at Section 6.3, Theorem 3 (non-amplification) and Definition 5 (A4)
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
Theorem 3 asserts that every admitted effect lies in the intersection of several operands, including the task/closure/state/purpose meet. The meet operation (Equation (2) and Appendix B) uses relational intersection of effect clauses. The paper notes that the executable kernel instantiates only exact-valued fields and does not implement general policy intersection or dominance elimination (Appendix B).
more of objection 3
This raises a concern: for a deployment with richer policy languages, the meet might not be representable, leading to a denial where a correct intersection exists. The paper states that 'a deployment using a richer language must retain the exact symbolic conjunction or fail closed'. This is prudent, but it effectively makes the Theorem depend on an implementation property that the present artifact does not realize. I recommend adding a brief remark in Section 6.3 or in the limitations (Section 12) that the non-amplification guarantee holds only when the meet is computable and the deployment can represent it exactly; otherwise the monitor may over-approximate (deny) but will not under-approximate (allow).
5 minor objections
The paper says, at Section 8.8
Pith says: minor objection 1 of 5
Table 5 reports 'All paths named' for 21 of 526 roots. The text says this is a 'strict lexical measure, not semantic dependency recall.' I suggest adding a one-sentence clarification that a root could still have complete closure even if it does not name every file lexically, if it uses a conservative enclosing container or a wildcard that the audit's lexical method does not recognize.
The paper says, at Section 5.3
Pith says: minor objection 2 of 5
The two-phase compilation rule (Figure 2) is described clearly, but the phrase 'cross-dimensional correlation is preserved' (end of Section 5.3) could benefit from a short example (e.g., the billing/refund counterexample). This exists earlier in the paragraph but a parenthetical or footnote would help.
show 3 more
The paper says, at Table 2
Pith says: minor objection 3 of 5
The table uses check marks and dashes. Consider adding a footnote that 'commit refresh' in the 'Prepared closure + effect' row means that the commit step rechecks effect and state but not closure or epochs; the naming could be clarified to avoid confusion with the full 'ClosureBound' profile.
The paper says, at Section 8.5
Pith says: minor objection 4 of 5
The sentence 'The 6/6 mutation score is targeted mechanism evidence ...' is a bit vague. Consider rephrasing to state explicitly that each of the six mutants produced the expected unsafe witness, confirming that the corresponding mechanism is necessary under the evaluated abstraction.
The paper says, at Section 7.1
Pith says: minor objection 5 of 5
The artifact is not included in the arXiv source bundle. I understand the retention policy, but the manuscript should mention that the internal verifier and machine-readable traces are available upon request to enable independent replication, even if not in the initial submission.
What would settle it
Select the 526 public Skill roots from the paper's own audit (Section 8.8). Implement the ClosureBound resolver (Algorithm 1) for these roots, using the same evidence profiles and canonicalization rules as the reference implementation. Count how many roots produce an 'exact resolvability' outcome per Definition 2 (all material nodes and edges have authenticated evidence, no floating versions, etc.). If fewer than 50% of roots satisfy the condition, then Assumption 1 is not met for typical current Skills, and the central guarantees cannot be claimed for that population without additional closure-discovery mechanisms.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 9 Sep 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee minor revision: 3 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Quick take: this is a solid design paper with an honest boundary. The contribution is the joint witness — binding the exact heterogeneous closure and the normalized realized effect in one operation-time authorization record — plus the non-inheritance and path-invariance framing. On the paper's own comparison against SIGIL, SkillGuard, AIRGuard, Cordon, and CommitGuard, that conjunction is new. I don't see a cited system doing it.
The best parts are the scoping discipline and the proofs. Every theorem states its assumptions and non-conclusions; Table 1 and Appendix D say plainly what is not established. The proofs are short but follow from the stated definitions; I did not find an internal contradiction or a load-bearing mathematical error. The external audit of 549 Skills is a genuinely useful piece of structural evidence, and the fact that it cuts against the completeness assumption — 21 of 526 roots name every path and none declares a dependencies field — is exactly why the authors include it. Credit for not burying that.
Soft spots, in proportion. First, the implementation, fixtures, and traces are withheld from the arXiv bundle, so the headline numbers (40/40, 6/6, 84,608 states) rest on self-report. That is addressable and not fatal, but an artifact release or an independent reimplementation would substantially change how much weight the evaluation deserves. Second, the safety invariants I1–I5 restate the admissibility conditions A1–A6 nearly verbatim, so the bounded exploration mostly verifies implementation-vs-spec consistency rather than an external property. The mutant witnesses show the check is not vacuous, but the ceiling is lower than the state count alone suggests. Third — and this is the real limitation — the theorems inherit Assumption 1: if the closure is incomplete, the guarantees are silent. The stress-test note is right that the paper's own audit shows current public Skills do not satisfy that assumption. I would not call this an internal contradiction, because the paper explicitly labels it a deployment obligation rather than a solved problem. But it is a clearly marked gap between the design and deployability, and it is not fixable by implementation effort alone; it needs a validated runtime discovery methodology and evaluation on real distributions.
The paper is for people building agent-skill runtime authorization, permission, or supply-chain systems. It is a useful reference design, and the limitation discussion is the best map I have seen of what remains to be done in this space. I would send it to a serious referee, with the expectation that the reviewer pushes hard on artifact availability and the completeness gap. The conditional core is coherent and worth engaging.
The strongest objection
The central claim (Theorems 2, 3, and 5) depends critically on Assumption 1: every material dependency appears as a node in the closure or is conservatively enclosed. The paper's own static audit (Section 8.8) provides strong evidence that current public Skills do not satisfy this condition: only 21 of 526 roots with bundled files name every path, and 67 contain links resolving outside their root. Furthermore, no root declares a traditional dependencies field. The paper labels this a 'deployment obligation' (Section 11.2) and admits that 'current distributions do not satisfy it' (Section 8.8). Without a validated method to achieve or approximate Assumption 1 in practice, the joint witness guarantee is vacuously true: if the closure is incomplete, the monitor may authorize an effect whose real dependencies were omitted, violating the non-inheritance and effect non-amplification claims. This is not an internal inconsistency—the paper is transparent about the dependency—but it is the single most load-bearing limitation because it applies to every real Skill and cannot be waived by implementation effort. The paper's mutants and fixtures all assume complete closure, so the evaluation does not address this concern.
The paper's formal results are logically sound given their assumptions, but the practical relevance hinges on Assumption 1. The reader identified this as the weakest assumption, and the paper's own audit provides external evidence that it fails in practice. This stress-test agrees with the reader's assessment and does not change the conditional verdict: the paper is a well-scoped design with honest limitations. The proposed concrete test would quantify the gap between assumption and reality, directly informing whether the approach is viable without additional dependency-discovery infrastructure.
The referee's report
If taken as a conditional design, the paper makes a crisp formal contribution: it separates exact transitive capability identity from root-level identity and produces a joint witness that closes several well-known authorization gaps in the Agent Skills ecosystem. The formalism is self-contained, the proofs are short but valid under the stated premises, and the paper is exemplary in scoping its claims (non-conclusions after every theorem, explicit trust summary in Appendix A). The external static audit (Section 8.8) provides independent structural evidence that current Skills do not satisfy the completeness assumption, which is valuable for motivating future work on dependency discovery. The paper does not claim a deployable product; it contributes a principled reference design that clarifies what a sound authorization decision must require.
Pith's reading of the paper (a machine's summary, not the authors')
Agent skills combine instructions, files, packages, tools, and services, so the operational capability can grow beyond a signed root directory. ClosureBound is a reference monitor that prevents authorization from being transferred to a different capability by re-resolving the entire typed dependency graph and normalizing the external effect at the point of commit. It admits an operation only if the current closure root, epochs, state, and effect all match the original grant bound. If the assumptions hold, no material change to any transitive dependency can inherit the old authority without a new explicit grant.
What carries the argument. The central object is the joint operation-time witness (Definition 5, Algorithm 2): an atomic predicate that checks the authenticated grant, the exactly re-resolved closure root (SCC-Merkle commitment), the currency of all epochs, the normalized effect falling in the intersection of the immutable grant ceiling and the current task/closure/state/purpose meet, and the freshness of a single-consumption reservation. This witness is evaluated at the commit linearization point, just before durability.
Core claim. The paper claims that a single joint operation-time witness, combining a trusted task grant, the exact Merkle-committed closure root, current runtime state, and a normalized effect IR, is both necessary and sufficient to prevent authorization transfer across changes to the heterogeneous dependency closure of an agent skill. It proves that under explicit assumptions, no material version change, epoch drift, effect-equivalent path switch, or stale reservation can be admitted without a new grant.
The premise it rests on. Assumption 1: Every material object that can influence a protected effect must appear as a node in the dependency closure, or be conservatively enclosed by a stronger pinned identity. If even one dependency path is missed—say, a runtime-loaded package or a URL that resolves to a changed service—the guarantee of non-inheritance collapses silently.
If the paper is right (Pith's inferences, not the authors')
Any material update to a transitive dependency, service deployment, or effect contract forces denial or explicit reauthorization, eliminating silent authority drift.
Supported equivalent-effect paths (e.g., direct API vs. browser automation) are checked against the same policy denotation, preventing path-based laundering.
Load-time integrity checks (hashes, signatures) become necessary but insufficient; only operation-time closure binding guarantees the current capability matches the grant.
Metadata and permissions declared in skill files cannot mint authority; only the authenticated task grant and independently resolved closure can authorize effects.
Replay attacks and stale pre-state are detected through the atomic reserve-and-consume step that rechecks every volatile fact at commit.
If ClosureBound is correct, it implies that the security of agent skill ecosystems depends on the completeness of dependency discovery—a deployment obligation that the paper's own audit suggests is unmet in current public distributions. This shifts the research bottleneck from authorization semantics to discovery instrumentation.
The approach could be extended to distributed systems by combining the operation-time witness with trusted execution environments or by requiring the remote service to consume the fence token, turning endpoint evidence into cryptographically enforced identity.
Because the grant is immutable and non-editable, ergonomics (automatic renewal vs. constant reauthorization) become a deployment policy choice: a prudent deployer may accept coarse-grained evidence for low-risk tasks while requiring full attestation for high-value effects.
The technique generalises beyond agent skills to any system where a composition of plugins, modules, or microservices can change after initial approval—e.g., VS Code extensions, CI/CD pipelines, or smart contract composability.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Agent Skills combine instructions with files, packages, tools, models, and services, so operational identity can exceed a signed directory. Recursive or lazy dependencies may change while root-level evidence remains valid, and different surfaces may reach the same durable effect. We present ClosureBound, a reference monitor that prevents authorization transfer across material changes to this heterogeneous closure. Its resolver commits typed graph nodes and topology. Each grant binds an exact closure root, effect ceiling, purpose/provenance, validity, and epochs. At durability, it re-resolves closure and state, normalizes the operation into an external-effect IR, and admits it only if a joint witness satisfies every bound. Supported equivalent paths share one ceiling. Assuming complete mediation and discovery, authenticated freshness, sound normalization, cryptographic binding, and authoritative linearization, we establish metadata non-authority, closure determinism, version non-inheritance, effect non-amplification, bound-value freshness, and path invariance. We do not establish program equivalence or remote-service honesty. A provider-free implementation matches 40 frozen lifecycle fixtures; 18 kernel contracts and six mutants cover binding and downgrade cases. Full-profile exploration reaches 84,608 states and 530,752 transitions without a declared invariant violation; six weakened profiles yield witnesses. A lexical audit of 549 public Skills (4,872 unique files) finds that 21 of 526 roots with bundled files name every non-manifest path verbatim, 67 contain links resolving outside their roots, and no root declares a frontmatter dependencies field. These observations motivate conservative closure discovery and define concrete targets for broader runtime, interoperability, efficacy, and production validation.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23
The authors' main claim, as the classifier picked it out 0.73
“Assuming complete mediation and discovery, authenticated freshness, sound normalization, cryptographic binding, and authoritative linearization, we establish metadata non-authority, closure determinism, version non-inheritance, effect non-amplification, bound-value freshness, and path invariance.”
How the main claim is stated: asserted 0.91
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.98 · human participants 0.03 · controlled comparison 0.08 · randomized assignment 0.01 · states quantitative result 0.98 · compares with prior results 0.12 · mentions code or data availability 0.26
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Agent Skills combine instructions with files, packages, tools, models, and services, so operational identity can exceed a signed directory. | background 1.00 | 0.08 | 0.04 | asserted 0.98 |
| Recursive or lazy dependencies may change while root-level evidence remains valid, and different surfaces may reach the same durable effect. | background 0.98 | 0.08 | 0.04 | hedged possibility 0.93 |
| We present ClosureBound, a reference monitor that prevents authorization transfer across material changes to this heterogeneous closure. | contribution 1.00 | 0.89 | 0.40 | asserted 1.00 |
| Its resolver commits typed graph nodes and topology. | method 0.92 | 0.77 | 0.04 | asserted 1.00 |
| Each grant binds an exact closure root, effect ceiling, purpose/provenance, validity, and epochs. | method 0.80 | 0.79 | 0.09 | asserted 1.00 |
| At durability, it re-resolves closure and state, normalizes the operation into an external-effect IR, and admits it only if a joint witness satisfies every bound. | method 0.99 | 0.79 | 0.10 | asserted 1.00 |
| Supported equivalent paths share one ceiling. | method 0.56 | 0.51 | 0.05 | asserted 1.00 |
| Assuming complete mediation and discovery, authenticated freshness, sound normalization, cryptographic binding, and authoritative linearization, we establish metadata non-authority, closure determinism, version non-inheritance, effect non-amplification, bound-value freshness, and path invariance. | result 0.56 | 0.74 | 0.45 | asserted 0.91 |
| We do not establish program equivalence or remote-service honesty. | limitation 1.00 | 0.81 | 0.06 | asserted 0.99 |
| A provider-free implementation matches 40 frozen lifecycle fixtures; 18 kernel contracts and six mutants cover binding and downgrade cases. | result 0.74 | 0.93 | 0.05 | asserted 1.00 |
| Full-profile exploration reaches 84,608 states and 530,752 transitions without a declared invariant violation; six weakened profiles yield witnesses. | result 0.98 | 0.94 | 0.07 | asserted 1.00 |
| A lexical audit of 549 public Skills (4,872 unique files) finds that 21 of 526 roots with bundled files name every non-manifest path verbatim, 67 contain links resolving outside their roots, and no root declares a frontmatter dependencies field. | result 0.98 | 0.94 | 0.04 | asserted 1.00 |
| These observations motivate conservative closure discovery and define concrete targets for broader runtime, interoperability, efficacy, and production validation. | future work 0.62 | 0.65 | 0.12 | asserted 0.90 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.40
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 9,527 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (3)
Figures are the authors', reproduced from the paper.
References (47)
- [1] Agent Skills Project. 2026. Adding Agent Skills Support to an Agent. https://agentskills.io/client- implementation/adding-skills-support. Accessed 2026-08-29 on Pith
- [2] Agent Skills Project. 2026. Agent Skills Specification. https://agentskills.io/specification . Accessed 2026-08-29; specification repository snapshot 69ef37e9424c0a7ea9dd2293b559e43ec8176379. Versioned Closure and Effect Governance for Agent Skills 29 on Pith
- [3] Anthropic. 2026. Skills: Public Repository for Agent Skills. https://github.com/anthropics/skills/tree/ 3b3fad96af16a10759d930941b4520ba0c40edae. Snapshot commit 3b3fad96af16a10759d930941b4520ba0c40edae, 2026-08-21; accessed 2026-08-29 on Pith
- [4] Sanket Badhe and Priyanka Tiwari. 2026. Agent Skill Security: Threat Models, Attacks, Defenses, and Evaluation. https://doi.org/10.48550/arXiv.2607.13987arXiv:2607.13987 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2607.13987
- [5] Lujo Bauer, Jay Ligatti, and David Walker. 2005. Composing Security Policies with Polymer. InProceedings of the 2005 ACM SIGPLAN Conference on Programming Language Design and Implementation. Association for Computing Machinery, New York, NY, USA, 305–314.https://doi.org/10.114... on PitharXiv
- [6] Varun Pratap Bhardwaj. 2026. Formal Analysis and Supply Chain Security for Agentic AI Skills. https://doi.org/ 10.48550/arXiv.2603.00195arXiv:2603.00195 [cs.CR] Version 2, revised 2026-08-05 on PithPith's reviewarXivdoi:10.48550/arxiv.2603.00195
- [7] Zheng Chen, Hanqing Liu, Duling Xu, Dong Dong, Jialin Li, Bangzheng Pu, and Jidong Zhai. 2026. Cordon: Semantic Transactions for Tool-Using LLM Agents.https://doi.org/10.48550/arXiv.2606.17573 arXiv:2606.17573 [cs.OS] on PithPith's reviewarXivdoi:10.48550/arxiv.2606.17573
- [8] Joseph W. Cutler, Craig Disselkoen, Aaron Eline, Shaobo He, Kyle Headley, Michael Hicks, Kesha Hietala, Eleftherios Ioannidis, John Kastner, Anwar Mamat, Darin McAdams, Matt McCutchen, Neha Rungta, Emina Torlak, and Andrew M. Wells. 2024. Cedar: A New Language for Expressive, ... on Pithdoi:10.1145/3649835
- [9] Jayanaka Dantanarayana, Savini Kashmira, Lingjia Tang, and Jason Mars. 2026. SIGIL: Compiling Agent Skills into Typed Harnesses. https://doi.org/10.48550/arXiv.2607.27309 arXiv:2607.27309 [cs.SE] Version 2, revised 2026-08-26 on PithPith's reviewarXivdoi:10.48550/arxiv.2607.27309
- [10] Wenbo Guo, Wei Zeng, Chengwei Liu, Xiaojun Jia, Yijia Xu, Lei Tang, Yong Fang, and Yang Liu. 2026. MalSkillBench: A Runtime-Verified Benchmark of Malicious Agent Skills. https://doi.org/10.48550/arXiv.2606.07131 arXiv:2606.07131 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2606.07131
- [11] Zimo Ji, Congying Xu, Zongjie Li, Yudong Gao, Xin Wei, Shuai Wang, and Shing-Chi Cheung. 2026. Cloak and Detonate: Scanner Evasion and Dynamic Detection of Agent Skill Malware. https://doi.org/10.48550/arXiv.2607.02357 arXiv:2607.02357 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2607.02357
- [12] Changguo Jia, Tianqi Zhao, Runzhi He, and Minghui Zhou. 2026. Skills Are Not Islands: Measuring Dependency and Risk in Agent Skill Supply Chains.https://doi.org/10.48550/arXiv.2607.01136arXiv:2607.01136 [cs.SE] on PithPith's reviewarXivdoi:10.48550/arxiv.2607.01136
- [13] Tu Lan and Chaowei Xiao. 2026. Runtime Skill Audit: Targeted Runtime Probing for Agent Skill Security. https: //doi.org/10.48550/arXiv.2606.11671arXiv:2606.11671 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2606.11671
- [14] Ying Li, Yanju Chen, Hongbo Wen, Bosi Zhang, Hanzhi Liu, Peiran Wang, Yu Feng, and Yuan Tian. 2026. VIGIL: Runtime Enforcement of Behavioral Specifications in AI Agent Skills. https://doi.org/10.48550/arXiv.2606.26524 arXiv:2606.26524 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2606.26524
- [15] Zhiyuan Li, Jingzheng Wu, Xiang Ling, Xing Cui, and Tianyue Luo. 2026. Towards Secure Agent Skills: Architecture, Threat Taxonomy, and Security Analysis. https://doi.org/10.48550/arXiv.2604.02837 arXiv:2604.02837 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2604.02837
- [16] Linux Foundation SPDX Project. 2025. SPDX Specification 3.0.1: Relationship. https://spdx.github.io/spdx- spec/v3.0.1/model/Core/Classes/Relationship/. Accessed 2026-07-25 on Pith
- [17] Xinyu Liu, Yukai Zhao, Xing Hu, and Xin Xia. 2026. Exploiting LLM Agent Supply Chains via Payload-less Skills. https://doi.org/10.48550/arXiv.2605.14460arXiv:2605.14460 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2605.14460
- [18] 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. https://doi.org/10.48550/arXiv. 2601.10338arXiv:2601.10338 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2601.10338
- [19] Microsoft. 2026. OpenAPM v0.1. https://microsoft.github.io/apm/specs/openapm-v01/ . Editor’s Working Draft; accessed 2026-07-25 on Pith
- [20] Model Context Protocol Project. 2026. Authorization and Security Best Practices. https://modelcontextprotocol. io/specification/2026-07-28/basic/authorization. Specification revision 2026-07-28; accessed 2026-08-29 on Pith
show 27 more
- [21] Model Context Protocol Project. 2026. Model Context Protocol 2026-07-28: Key Changes. https : / / modelcontextprotocol.io/specification/2026-07-28/changelog. Accessed 2026-08-29 on Pith
- [22] Bardia Mohammadi, Nearchos Potamitis, Lars Klein, Akhil Arora, and Laurent Bindschaedler. 2026. Atomix: Timely, Transactional Tool Use for Reliable Agentic Workflows. https://doi.org/10.48550/arXiv.2602.14849 arXiv:2602.14849 [cs.LG] on PithPith's reviewarXivdoi:10.48550/arxiv.2602.14849
- [23] OpenAI. 2026. OpenAI Plugins: Curated Codex Plugin Examples. https://github.com/openai/plugins/tree/ 1e285826e604f66f7208f7ac4dba0fe8341d1f57. Snapshot commit 1e285826e604f66f7208f7ac4dba0fe8341d1f57, 2026-08-28; accessed 2026-08-29 on Pith
- [24] OpenSSF SLSA Community. 2026. SLSA Build: Verifying Artifacts. https://slsa.dev/spec/v1.2/verifying- artifacts. Accessed 2026-07-25 on Pith
- [25] OpenSSF SLSA Community. 2026. SLSA Specification, Version 1.2. https://slsa.dev/spec/v1.2/ . Accessed 2026-07-25. 30 Zhu and Wang on Pith
- [26] OWASP Foundation. 2025. CycloneDX Bill of Materials Standard. https://cyclonedx.org/specification/ overview/. Accessed 2026-07-25 on Pith
- [27] Nils Palumbo, Sarthak Choudhary, Jihye Choi, Guy Amir, Prasad Chalasani, and Somesh Jha. 2026. Formal Policy Enforcement for Real-World Agentic Systems. https://doi.org/10.48550/arXiv.2602.16708 arXiv:2602.16708 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2602.16708
- [28] Shidong Pan, Xiaoyu Sun, Tianyi Zhang, Dianshu Liao, Kaiwen Yang, and Zhenchang Xing. 2026. SkillGuard: A Permission-Centric Framework for Agent Skill Security. https://doi.org/10.48550/arXiv.2606.03024 arXiv:2606.03024 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2606.03024
- [29] Jaehong Park and Ravi Sandhu. 2004. The UCONABC Usage Control Model.ACM Transactions on Information and System Security7, 1 (2004), 128–174.https://doi.org/10.1145/984334.984339 on PitharXiv
- [30] Suliu Qin, Haomin Zhuang, Yujun Zhou, Yufei Han, and Xiangliang Zhang. 2026. AIRGuard: Guarding Agent Actions with Runtime Authority Control.https://doi.org/10.48550/arXiv.2605.28914arXiv:2605.28914 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2605.28914
- [31] Shoumik Saha, Kazem Faghih, and Soheil Feizi. 2026. Under the Hood of SKILL.md: Semantic Supply-Chain Attacks on AI Agent Skill Registry.https://doi.org/10.48550/arXiv.2605.11418arXiv:2605.11418 [cs.AI] on PithPith's reviewarXivdoi:10.48550/arxiv.2605.11418
- [32] Jerome H. Saltzer and Michael D. Schroeder. 1975. The Protection of Information in Computer Systems.Proc. IEEE63, 9 (1975), 1278–1308.https://doi.org/10.1109/PROC.1975.9939 on PitharXiv
- [33] Igor Santos-Grueiro. 2026. Temporary Authority, Permanent Effects: Commit-Time Authorization for LLM Agents. https://doi.org/10.48550/arXiv.2607.10487arXiv:2607.10487 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2607.10487
- [34] Fred B. Schneider. 2000. Enforceable Security Policies.ACM Transactions on Information and System Security3, 1 (2000), 30–50.https://doi.org/10.1145/353323.353382 on PitharXiv
- [35] Tingda Shen, Yebo Feng, Konglin Zhu, Xiaojun Jia, Yang Liu, and Lin Zhang. 2026. Sealing the Audit-Runtime Gap for LLM Skills.https://doi.org/10.48550/arXiv.2605.05274arXiv:2605.05274 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2605.05274
- [36] The Update Framework Project. 2026. The Update Framework Specification. https://theupdateframework.github. io/specification/latest/. Accessed 2026-07-25 on Pith
- [37] Guiyao Tie, Jiawen Shi, Pan Zhou, and Lichao Sun. 2026. BadSkill: Backdoor Attacks on Agent Skills via Model-in-Skill Poisoning.https://doi.org/10.48550/arXiv.2604.09378arXiv:2604.09378 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2604.09378
- [38] Santiago Torres-Arias, Hammad Afzali, Trishank Karthik Kuppusamy, Reza Curtmola, and Justin Cappos. 2019. in-toto: Providing Farm-to-Table Guarantees for Bits and Bytes. In28th USENIX Security Symposium. USENIX Association, Santa Clara, CA, 1393–1410. https://www.usenix.org/co... on Pith
- [39] Jiangrong Wu, Yuhong Nan, Yixi Lin, Huaijin Wang, Yuming Xiao, Shuai Wang, and Zibin Zheng. 2026. SkillScope: Toward Fine-Grained Least-Privilege Enforcement for Agent Skills. https://doi.org/10.48550/arXiv.2605. 05868arXiv:2605.05868 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2605
- [40] Yi Xie, Jiawei Du, Yu Cheng, Jiuan Zhou, and Zhaoxia Yin. 2026. Benign in Isolation, Harmful in Composition: Security Risks in Agent Skill Ecosystems.https://doi.org/10.48550/arXiv.2606.15242arXiv:2606.15242 [cs.CR] on Pithdoi:10.48550/arxiv.2606.15242
- [41] Zhonghao Zhan and Hamed Haddadi. 2026. Auto-Policy, not Auto-Skill: Compiled Agent Skills for the Physical World. https://doi.org/10.48550/arXiv.2608.25091arXiv:2608.25091 [cs.AI] on PithPith's reviewarXivdoi:10.48550/arxiv.2608.25091
- [42] Ziling Zhou. 2026. Governing Dynamic Capabilities: Cryptographic Binding and Reproducibility Verification for AI Agent Tool Use.https://doi.org/10.48550/arXiv.2603.14332arXiv:2603.14332 [cs.CR] on Pithdoi:10.48550/arxiv.2603.14332
- [43] Genliang Zhu and Chu Wang. 2026. Explanation-Bound Tool Execution for AI Agents: Server-Verified Action Claims Without Trusting Model Rationales.https://doi.org/10.48550/arXiv.2607.25364arXiv:2607.25364 [cs.AI] on PithPith's reviewarXivdoi:10.48550/arxiv.2607.25364
- [44] Genliang Zhu and Chu Wang. 2026. Intent-Governed Tool Authorization for AI Agents. https://doi.org/10. 48550/arXiv.2606.22916arXiv:2606.22916 [cs.AI] on PithPith's reviewarXivdoi:10.48550/arxiv.2606.22916
- [45] Genliang Zhu, Chu Wang, Ziyuan Wang, Zhida Li, and Qiang Li. 2026. OpenPort Protocol: A Security Governance Specification for AI Agent Tool Access. https://doi.org/10.48550/arXiv.2602.20196 arXiv:2602.20196 [cs.CR] on Pithdoi:10.48550/arxiv.2602.20196
- [46] Haomin Zhuang, Hanwen Xing, Yujun Zhou, Yuchen Ma, Yue Huang, Yili Shen, Yufei Han, and Xiangliang Zhang on Pith
- [2026] AgentTrap: Measuring Runtime Trust Failures in Third-Party Agent Skills. https://doi.org/10.48550/ arXiv.2605.13940arXiv:2605.13940 [cs.CR] on PithPith's reviewarXivdoi:10.48550/arxiv.2605.13940
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment