← MCP server and skill grades

skills

Yuan1z0825/nature-skills/skills

D
Type
MCP server
Scanned
Nov 13, 2026
Version
7bff1d0539c0
Files read
758

This toolkit gives AI assistants skills to search academic databases, download research papers, generate figures, and turn manuscripts into Word documents, PowerPoint slides, or patent drafts. The biggest risks are that malicious Word templates or paper downloads can read private files or save files outside the download folder on Windows, and the toolkit exposes your logged-in Chrome browser over an unprotected local connection while installing unverified software packages without asking.

What it can reach

Files
Reads and writes project files (including manuscripts, PDFs, Word and PowerPoint documents, images, notes, Zotero data, and codebases) along with settings and cache folders in your home directory (such as ~/.config, ~/.claude, ~/.image2ppt, and ~/.cache).
Network
Connects to academic and patent databases (such as PubMed, arXiv, CrossRef, Scopus, Web of Science, CNKI, CNIPA, and Google Scholar), cloud AI and text-recognition services (including OpenAI/ChatGPT, OpenRouter, Baidu PaddleOCR, and Firecrawl), and messaging apps like Feishu and Telegram.
Secrets
Reads and stores API keys and login tokens from environment variables and local config files, including publisher keys (Elsevier, Springer Nature, IEEE, NCBI), AI and OCR keys (OpenAI, OpenRouter, Baidu PaddleOCR, Firecrawl), and your Codex login token in ~/.codex/auth.json.
Browser
Controls your live, logged-in Chrome browser via a local remote-debugging proxy (127.0.0.1:3456) to browse pages, run scripts, and download papers using your saved logins, and launches automated Chromium browsers for patent searches and diagram rendering.
Shell commands
Runs local command-line programs and scripts (including Python, Node.js, R, PowerShell, LaTeX, ImageMagick, LibreOffice, and lark-cli) and uses package managers like pip, npm, npx, uv, and Playwright to install and run tools.
System
Can schedule recurring background tasks (cron jobs) for daily paper summaries and automate desktop Microsoft PowerPoint or LibreOffice to convert and check slides.

Findings

2 high2 medium6 low
  1. HighXML external entity (XXE) local file disclosure when inspecting reference Word documents

    skills/nature-paper2word/scripts/reference_layout.py:20

    When reformatting a manuscript to match a supplied reference .docx file, reference_layout.py extracts word/document.xml from the archive and parses it with lxml.etree.fromstring() using lxml's default parser. Because lxml enables resolve_entities=True by default and allows file:// SYSTEM entities even when network entities are disabled, a malicious .docx template or reference paper can embed local files from the user's computer into the extracted paragraph text written to reference-samples.json and read by the AI agent, or trigger exponential entity expansion.

    Fix: Parse OOXML parts with defusedxml or pass lxml.etree.XMLParser(resolve_entities=False, no_network=True) to etree.fromstring().

  2. HighWindows path traversal via unsanitized backslashes in downloaded filenames

    skills/nature-downloader/scripts/batch_download.mjs:339

    Filename sanitizers in batch_download.mjs (downloadDoi and downloadSi) as well as scripts/lib/direct-download.mjs, scripts/lib/cnki.mjs, and scripts/lib/open-access.mjs use the character class /[\/:*?"<>|]/, which strips forward slashes (/) but omits backslashes (\). On Windows, backslashes remain valid path separators and are preserved verbatim in URL query strings such as ?file=\..\..\... on a supplementary link or inside Crossref title metadata, allowing untrusted remote metadata or webpage links to write or overwrite arbitrary files outside the target --out directory.

    Fix: Include \\ in all filename sanitization regular expressions, strip leading/trailing dots, apply path.basename(), and verify that resolved output paths remain inside the intended output directory before writing.

  3. MediumServer-side request forgery via unvalidated open-access metadata URLs

    skills/nature-downloader/scripts/lib/open-access-provider.mjs:28

    During open-access lookup, downloadOpenAccessArticle fetches candidate URLs supplied by Crossref link records and Unpaywall responses directly from Node.js without checking the URL scheme, hostname, resolved IP address, or HTTP redirects. A crafted DOI metadata record with an open license can point to localhost services (such as the unauthenticated local CDP proxy on 127.0.0.1:3456), cloud metadata endpoints, or private intranet hosts, saving matching PDF/XML/HTML responses into the download folder and recording HTTP status codes in manifest.json.

    Fix: Validate that candidate URLs and any redirect targets use HTTPS and resolve only to public global IP addresses before fetching.

  4. MediumSilent runtime download and execution of unpinned npm package via npx -y

    skills/nature-paper-to-patent/scripts/disclosure/mermaid_render.py:67

    When rendering Mermaid diagrams for patent disclosures and mmdc is not already installed locally or globally, mermaid_render.py falls back to running npx -y @mermaid-js/mermaid-cli mmdc. The -y flag suppresses confirmation and automatically downloads and executes the latest unpinned version of @mermaid-js/mermaid-cli and its transitive dependencies from npm at runtime.

    Fix: Require pre-installation from a lockfile with pinned versions instead of invoking npx -y with an unpinned package specifier at runtime.

  5. LowCross-origin Same-Origin-Policy bypass via authenticated browser navigation fallback

    skills/nature-downloader/scripts/lib/wos-supporting-information.mjs:111

    When downloading Supporting Information in the Web of Science bundle workflow, links labelled as supplementary material on a DOI landing page are first fetched from that page's origin, and only five hardcoded repository domains are excluded. If Chrome blocks a cross-origin link due to CORS, fetchAttachmentWithNavigationFallback catches the CORS error, navigates the user's logged-in Chrome tab directly to that cross-origin URL, and re-fetches the resource from the target origin with credentials included, saving any non-HTML file from other authenticated sites to disk.

    Fix: Restrict navigation-fallback downloads to the same registrable domain or verified publisher asset hosts rather than navigating the authenticated session to arbitrary cross-origin URLs.

  6. LowPlaywright Chromium launched with --no-sandbox over unencrypted HTTP

    skills/nature-paper-to-patent/scripts/disclosure/cnipa_epub_crawler.py:195

    The CNIPA prior-art search helper unconditionally launches Chromium with --no-sandbox and navigates to http://epub.cnipa.gov.cn/ over plaintext HTTP, executing page JavaScript for up to 180 seconds. On an untrusted network (such as public or campus Wi-Fi), an active man-in-the-middle attacker can inject arbitrary JavaScript into a browser renderer whose OS sandbox is disabled, and can tamper with the patent search results returned to the AI agent.

    Fix: Remove --no-sandbox by default (or gate it behind an explicit container-only environment opt-in) and use HTTPS if supported by the endpoint.

  7. LowSkill instructions direct the agent to install an unpinned PyPI package at runtime

    skills/nature-paper2word/SKILL.md:15

    Before exporting a manuscript, the skill instructs the AI agent to run pip install md-to-word-docx automatically if the md-to-word command is missing. Installing an unpinned package without version or hash constraints during normal skill execution exposes the user to upstream PyPI supply-chain changes without explicit review.

    Fix: Pin md-to-word-docx to a vetted version in a requirements.txt file and ask for user confirmation before installing packages.

  8. LowLaTeX formula renderer lacks -no-shell-escape and file-inclusion restrictions

    skills/nature-image2ppt/cli/image2ppt/runtime/formula_renderer.py:78

    Although nature-image2ppt states that page operations may read and write only inside their assigned page directory, render_latex_asset invokes xelatex, lualatex, or pdflatex without -no-shell-escape (leaving restricted write18 and LuaTeX code execution active), without restricting openin_any (allowing \input or \openin to render arbitrary local files into slide formula images), and exposes a --shell-escape CLI flag. If prompt injection in a scanned slide tricks the agent into transcribing malicious TeX or flags, the page directory boundary is bypassed.

    Fix: Always pass -no-shell-escape (and --safer for lualatex), set openin_any=p and openout_any=p in the TeX environment, and remove the --shell-escape CLI option.

  9. LowDOM-based cross-site scripting in generated citation review HTML

    skills/nature-citation/scripts/nature_citation.py:2175

    When nature_citation.py is run with --with-artifacts, the generated citation.html embeds raw bibliographic strings in JSON and injects item.title and item.journal into wrapper.innerHTML when a user checks a reference checkbox. Because journal names fetched by DOI are not stripped of HTML tags and clean_text() on titles does not strip unclosed tags such as <img src=x onerror=..., crafted metadata from an external source can execute JavaScript when the user opens the local HTML review page and selects the reference.

    Fix: Construct the h4 and p elements with document.createElement() and assign item.title, item.journal, and item.year via textContent instead of innerHTML.

  10. LowUnpinned npx -y fallback downloads and runs @mermaid-js/mermaid-cli at runtime without confirmation

    skills/nature-paper-to-patent/references/disclosure/tooling.md:47

    When a local mmdc binary has not been preinstalled (the default state of a freshly installed skill), the technical-disclosure workflow and troubleshooting instructions invoke npx -y @mermaid-js/mermaid-cli mmdc. Passing -y without a pinned version silently downloads and executes whatever version is currently tagged latest on npm, creating a supply-chain code-execution risk.

    Fix: Pin @mermaid-js/mermaid-cli to an exact reviewed version, remove the automatic -y fallback so missing dependencies fail closed with setup instructions, and ask the user before installing Node packages.