acceptodds
Under review as a conference paper at ICLR 2027

Copied, Fetched, and Left Without Evidence: Permissive Washing in AI-Agent Supply Chains

Abstract

Permissive washing occurs when an artifact declares MIT, Apache-2.0, or BSD-3-Clause without the matching license text and required copyright holders that form its compliance payload. Documented in datasets and models, it now affects an AI-agent ecosystem built on shared Skills, Model Context Protocol (MCP) servers, and Plugins. These components are reused and distributed like software. Authors copy folders and files into repositories, and installers and registries deliver selected directories and packages. Their permissive licenses condition such copying and redistribution on retaining notices, so losing the compliance payload can have legal consequences. We trace the compliance payload across publication, reuse, composition, and distribution. Among 345,371 assessed Skill occurrences, 59.4% lack a complete compliance payload anywhere from folder to repository root, and only 9.5% carry one inside the Skill folder. Even when the reference was complete before reuse, 41.1% of Skill file copies and 58.5% of MCP server file copies (copies made without forking) lose it. Distribution loses it the same way. Claude Code leaves 173 of 229 repository-complete Skill deliveries without either part, Codex agrees on every shared outcome, and 22 of 121 registry archives for MCP servers omit the compliance payload. The compliance payload survives when the unit copied or delivered encloses it, not merely when the repository does. We release all data and code in a replication package.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, AI Agents, AI Supply Chains, License Compliance, AI infrastructure

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MamHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper audits whether permissively licensed agent components retain their license text and required copyright notices across publication, reuse, composition, and distribution. It studies Skills, MCP servers, and Plugins using ScanCode-based payload detection, content matching, provenance and timing filters, and installation and registry-archive checks. The authors assess 345,371 Skill occurrences and report that 59.4% lack a complete payload along their governing directory chains. Among time-eligible copies with complete references, they report payload loss in 41.1% of Skill file copies and 58.5% of MCP server file copies. Controlled installations with Claude Code and Codex and inspection of 121 registry archives examine whether repository-level notices survive delivery. The central finding is that complete repository documentation does not ensure complete component copies: preservation depends on whether the copied or delivered unit encloses the relevant notices. The authors also describe a replication package of data, code, and audit artifacts.

Critical strengths

S1: The paper moves beyond checking repository license labels to tracing the material retained in component copies, bundled Plugins, installed directories, and registry archives. The distinction between a complete repository and a complete delivered component is relevant to publishers and distribution tooling, and the installation and archive checks make the contribution more concrete than a metadata-only audit.

S2: The authors exclude copies predating their reference's complete payload, examine README notices and repository weighting, disclose concentration in large collections, and compare two installers. The byte-preservation checks also help locate the measured loss at directory selection. These controls address important alternative explanations and deserve credit.

Critical weaknesses

W1: The C4 delivery boundary does not yet establish the stronger recipient-level interpretation. Section 3 compares the full repository, the selected directory, and the installed cache (pp. 5, 7–8). This establishes whether the selected plugin copy is self-contained, but Section 2.2 also recognizes that accompanying documentation can carry notices. Current Claude Code loading documentation distinguishes fetched marketplace copies from plugin-cache copies, suggesting a concrete alternative: an ancestor license could reach the recipient through an automatically fetched marketplace checkout while remaining absent from the plugin cache. This must be checked for the paper's pinned versions and source types, current documentation alone does not invalidate its historical results. Please distinguish pre-seeded source snapshots from automatically delivered artifacts and show whether required notices are absent from the complete acquisition flow. Otherwise, the result should be framed as a lack of self-contained notices in the selected copy, with a separate explanation of its implications for recipients and subsequent redistribution. This affects the interpretation of a central contribution beyond existing repository-level audits.

W2: The provenance-to-required-notice mapping needs more direct validation. Section 3 uses the earliest dated matching Skill body as a reference and tests preservation of its required holders (p. 4). Temporal precedence and content equality do not by themselves establish the relevant origin or distinguish a shared upstream source. Likewise, notices found in ancestor directories need to be connected to the particular copied component. The timing gate addresses postdated payloads, and the reported 83.4% MCP pair precision validates candidate matching, but neither alone validates the final attribution-loss labels. This matters because every failed C2 Skill copy is missing a required holder, and 79 of the 89 failed C3 Skill copies retain matching license text (pp. 6–7). Please provide the component-to-notice applicability rules and existing or bounded manual validation of the final labels, including sensitivity to uncertain provenance or unrelated notices. This concern does not negate C1's widespread missing license text; it affects the stronger attribution-tracing conclusions. Relevant validation may be in the appendices unavailable to this assessment.

Questions for authors

Q1: For the pinned Claude Code/Codex versions, what exact files are automatically acquired during marketplace registration and plugin installation, beyond the final plugin cache? Do representative failed C4 cases retain the required notices in an automatically fetched marketplace copy or accompanying documentation? Please separate these from manually pre-seeded repositories and report cache-level versus complete-acquisition outcomes, or identify the existing supplementary evidence that establishes this distinction.

Q2: How are notices in a reference's governing chain determined to apply to the matched component, and how are common upstream sources, unrelated notices, and ambiguous holder matches handled? Please identify any existing manual audit of the final C2/C3 attribution-loss labels and its results, or provide a bounded stratified check and show how excluding uncertain cases changes the reported outcomes. Candidate-pair precision should be distinguished from the precision of the final loss labels.

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

This paper studies whether permissively licensed AI-agent components preserve their license text and copyright holders during publication, reuse, composition, and distribution. The authors conduct a large-scale analysis of Skills, MCP servers, and Plugins and find that license information is frequently missing or lost during copying and installation. The results show that having complete license information at the repository level does not guarantee that it will be preserved in the components actually delivered to users.

Critical strengths
  1. The paper provides a large-scale analysis of license compliance across Skills, MCP servers, and Plugins, covering four stages of the AI-agent supply chain: publication, reuse, composition, and distribution.

  2. The paper identifies an important issue where license information may be lost when components are copied or installed, even if the original repository contains complete license information.

Critical weaknesses
  1. The attribution checks need stronger validation. The paper identifies Skill copies by matching their contents and treats the earliest occurrence as the reference. However, this does not necessarily mean that the copyright holders in the reference repository should also appear in the later copy. Both copies could come from another source, or the reference repository's license may cover other files rather than the specific Skill. Since all 713 failed Skill copies are missing at least one required copyright holder, this assumption may affect the main results.

  2. The generalizability of the distribution results is limited. As the authors acknowledge, 168 of the 173 Claude Code losses come from only two Skill collections. Although the paper also provides an exposure analysis, this mainly measures where license files are stored rather than whether they are actually lost during installation. Therefore, it remains unclear how common this problem is across the broader AI-agent ecosystem. Reporting results across more independent repositories or Skill collections would strengthen the findings.

  3. The novelty of the work is somewhat limited. Missing license information during software copying and distribution has already been studied in traditional software supply chains. The paper mainly applies existing license detection tools and analysis methods to AI-agent components. Although the study provides useful empirical results, it remains unclear what new technical insights or methods are introduced beyond extending previous license compliance studies to Skills, MCP servers, and Plugins.

Questions for authors

Please refer to the weaknesses above. In particular, I would appreciate clarification on how copyright holders are attributed to copied Skills, how representative the observed distribution failures are, and what distinguishes this study from existing software license compliance research.

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

The paper audits whether license text and copyright-holder notices remain with AI-agent Skills, MCP servers, and Plugins as they are published, copied, bundled, and delivered. It reports missing machine-detectable license material at each stage, including losses when installers select a subdirectory but leave the repository’s license files behind. The evidence is a large crawl and a smaller set of copy, bundling, installation, and package checks.

Critical strengths
  • The underlying problem is real and current. OpenAI’s docs describe Skills as directories of instructions and resources and Plugins as installable packages; Anthropic’s docs likewise describe marketplace plugins as directories of components. These distribution patterns make it plausible for a selected component to be separated from repository-level notices. The public docs I checked do not establish that license checks are absent from every product, but they do not document a safeguard that resolves the paper’s specific selected-directory problem. The paper identifies a concrete failure mode and tests actual installed outputs, not only repository metadata.
Critical weaknesses
  • The paper moves too quickly from “our scanner did not find the full payload in this unit” to “permissive washing” and possible legal noncompliance. The license terms differ in what must accompany a redistributed copy: for example, BSD-3-Clause permits notices in documentation or other distribution materials for binaries, and Apache 2.0 allows certain attribution notices in source, documentation, or generated displays. Whether a short Skill is a substantial portion, whether files elsewhere accompany the delivered unit, and whether the source and destination have the same owner also matter. The results should consistently be presented as missing machine-detectable evidence under the study’s rule, not as adjudicated legal violations.
  • The headline rates depend on provenance and detection rules whose validation is not assessable from the provided PDF. Skill copies are attributed by matching bodies to the earliest dated occurrence, which can confuse a shared template or independent reuse with a direct copy. The reported MCP copy rule has 83.4% precision, but the main text does not give recall or the loss rate among verified pairs; the license/holder detector’s manual accuracy is also not reported there. In addition, two collections account for 168 of the 173 Claude Code Skill losses, and the cited Appendices A–N are not present in the available submission PDF. The large counts and installer demonstration are useful, but they do not yet support broad ecosystem or legal conclusions.
Questions for authors
  1. Have you checked whether current frontier-provider marketplaces or installers enforce license-retention checks? The public docs establish that Skills and Plugins are distributed, but do not show whether these safeguards exist. Please distinguish observed delivery losses from inferences based only on product documentation.
  2. Please provide the cited appendices and report manual precision/recall for payload detection and copy attribution, including the MCP loss rate on verified copy pairs and the C4 results after removing the two dominant collections. These checks would materially affect my confidence in the prevalence claims.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.