Back to skill

Security audit

Paper to Slides

Security checks for vulnerabilities and agentic risk

Overview

The skill mostly does what it claims, but it asks the agent to fetch arbitrary PDF URLs with raw curl and weak safeguards, which deserves manual review before installation.

Install only if you are comfortable with the agent reading user-supplied PDFs, downloading remote papers, and generating HTML that may load third-party fonts. Prefer local trusted PDFs or trusted academic URLs, avoid sensitive internal URLs, and consider hardening the workflow to use per-run private temp directories and a downloader that validates HTTPS, destination, redirects, size, and PDF content before extraction.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (2)

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:26
Finding

Unrestricted User-Controlled URL Retrieval Enables Server-Side Request Forgery

Content
View full analysis
" ``` ``` ### Technical Analysis The Skill instructs the Agent to pass a user-provided PDF URL directly to `curl`. Although the URL is enclosed in shell quotes, which reduces conventional shell-command injection risk, no validation is applied to: - The URL scheme - The destination hostname or resolved IP address - The destination port - Redirect targets - The response size or download duration - The response content type or PDF file signature The `-L` option causes `curl` to follow redirects. Consequently, even if an initial URL points to a public host, that host can redirect the Agent to a loopback, private-network, link-local, or cloud metadata address. Depending on the installed `curl` build and accepted input, non-HTTP schemes may also be reachable unless explicitly prohibited. Because the request originates from the Agent host, it carries that host's network reachability rather than the user's network reachability. ### Attack Path 1. An attacker submits a purported PDF URL that targets an internal service, loopback endpoint, link-local metadata service, or an attacker-controlled public redirector. 2. The Skill executes: ```bash curl -L -o /tmp/paper.pdf "" ``` 3. `curl` connects from the Agent environment and follows any supplied redirect chain. 4. The internal or otherwise restricted response is written to `/tmp/paper.pdf`. 5. The Skill subsequently processes the downloaded response with PDF extraction or other analysis tools. 6. Depending on response handling, the attacker may infer service availability or cause retrieved content to be included in generated reports and slides. ### Impact ...[truncated 575 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Note
Location
SKILL.md:30
Finding

Predictable Temporary File Paths Permit Symlink Clobbering and Cross-Run Data Contamination

Content
View full analysis
" ``` ```bash pdftotext "" /tmp/paper_text.txt ``` ```markdown Create `temp_analysis.md` extracting: - Research question, hypotheses, methodology, data sources - Core findings with key quantitative results - Theoretical contributions and practical implications - The paper's fundamental contradiction/gap, novel angle, and method innovation ``` ### Technical Analysis The workflow uses fixed, predictable intermediate filenames, including: - `/tmp/paper.pdf` - `/tmp/paper_text.txt` - `temp_analysis.md` in an unspecified working directory Shared temporary directories are commonly writable by multiple local users. If the invoked tools follow symbolic links when opening output files, another local process can pre-create a symbolic link at one of these predictable paths. A subsequent Skill run may then overwrite the symlink target with the privileges of the Agent process. Fixed filenames also create race conditions between concurrent Skill invocations. One run can overwrite another run's paper or extracted text, causing incorrect analysis, accidental disclosure, or attacker-controlled substitution of content. No private per-run temporary directory, exclusive file creation, ownership verification, restrictive permissions, or cleanup procedure is specified. ### Attack Path 1. A local attacker predicts that the Skill will use `/tmp/paper.pdf` or `/tmp/paper_text.txt`. 2. Before the Skill writes the file, the attacker creates a symbolic link at that path pointing to another file writable by the Agent. 3. The Skill executes `curl` or `pdftotext` using the predictable output path. 4. If the output-opening behavior follows the symbolic link, data is written to the unintended target. 5. Alternatively, an attacker or concurrent Skill run repl ...[truncated 733 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (16)

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · prompts/slide-template.md (reported line 15)May include surrounding context.

md
<meta name="viewport" content="width=device-width, initial-scale=1.0">
    <title>Presentation Title</title>

    <!-- Fonts from Google Fonts or Fontshare — never system fonts -->
    <link href="https://fonts.googleapis.com/css2?family=..." rel="stylesheet">

    <style>

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · style-preview.html (reported line 272)May include surrounding context.

html
<div class="progress-bar" id="progressBar"></div>
    <nav class="nav-dots" id="navDots"></nav>

    <!-- 1. Swiss Modern -->
    <section class="slide swiss-modern">
        <span class="style-label reveal">Swiss Modern</span>
        <span class="style-number reveal">01 / 10</span>

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The README advertises trigger phrases like 'read this paper and make slides' and similar broad variants, which overlap with normal user requests. In agent ecosystems that auto-route by fuzzy matching, this can cause the skill to activate unexpectedly on generic prompts and then fetch/process local PDFs or remote URLs without sufficiently explicit user intent. The risk is increased here because the skill supports both local file paths and network retrieval of papers.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The README explicitly says the skill accepts arXiv links and arbitrary PDF URLs and later notes that network access and Google Fonts may be used, but it does not clearly warn users up front that the skill will fetch remote content and load external resources. In an agent setting, silent network retrieval can expose sensitive prompts, metadata, or user intent to third parties and can violate offline or restricted-network expectations.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · README.md (reported line 84)May include surrounding context.

brew install poppler

Ubuntu/Debian

sudo apt install poppler-utils

text

完成。不需要 Python 包、npm 或构建工具。生成的 slides 是纯 HTML + 内联 CSS/JS。

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · README_zh.md (reported line 84)May include surrounding context.

brew install poppler

Ubuntu/Debian

sudo apt install poppler-utils

text

完成。不需要 Python 包、npm 或构建工具。生成的 slides 是纯 HTML + 内联 CSS/JS。

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The trigger scope is broad enough to match many ordinary summarization or presentation requests, which can cause the agent to invoke this skill unexpectedly. Over-broad activation increases the chance of unnecessary file handling, network access, and tool execution outside the user's intended workflow.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill instructs the agent to download remote PDFs from arbitrary URLs without warning the user that network access will occur. This creates privacy and trust risks because simply invoking the skill may contact external hosts, disclose interest patterns or metadata, and retrieve untrusted content into local temporary storage.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The instructions mandate bilingual terms on first mention, and the slide parameter prompts are presented only in Chinese, which imposes a language/locale preference without asking the user. This is a natural-language policy concern because the skill does not offer language choice or document a justified locale restriction.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The template title and instructions require the response to follow a Chinese-language structure and wording, but there is no indication that the user can opt into another language. This is a natural-language locale policy issue because it imposes a specific language by default rather than offering a choice or documenting a region-specific justification.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The HTML template hard-codes lang="zh-CN", which imposes a specific language/locale on all generated presentations. The file does not offer user opt-in or explain that the template is intended only for a China-specific context, so this is a natural-language locale policy issue.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The document declares lang="zh-CN", and the visible content throughout the page is presented in Chinese without any indication that users can choose another language. This is a natural-language locale constraint that may violate organizational language-choice policy when not explicitly opt-in or justified as region-specific.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

The page loads third-party font resources from Google Fonts and Fontshare, causing user browsers to contact external domains and disclose metadata such as IP address, user agent, timing, and referrer context. In a local/offline skill preview, this creates unnecessary external data transmission and introduces a supply-chain/dependency surface if those resources are modified or unavailable.

Content

Scanner excerpt · style-preview.html (reported line 11)May include surrounding context.

html
<link rel="preconnect" href="https://fonts.googleapis.com">
    <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
    <link href="https://fonts.googleapis.com/css2?family=Archivo:wght@400;700;800;900&family=Archivo+Black&family=Nunito:wght@400;500;600&family=Space+Grotesk:wght@400;500&family=Cormorant:ital,wght@0,400;0,600;1,400&family=IBM+Plex+Sans:wght@300;400&family=Bodoni+Moda:wght@400;700&family=DM+Sans:wght@400;500&family=Plus+Jakarta+Sans:wght@400;500;700;800&family=Outfit:wght@400;500;700;800&family=Fraunces:wght@700;900&family=Work+Sans:wght@400;500&family=JetBrains+Mono:wght@400;500&family=Cormorant+Garamond:wght@400;600&family=Source+Serif+4:wght@400;600&family=Space+Mono:wght@400;700&display=swap" rel="stylesheet">
    <link rel="stylesheet" href="https://api.fontshare.com/v2/css?f[]=clash-display@700&f[]=satoshi@400,500&display=swap">

    <style>
        * { margin: 0; padding: 0; box-sizing: border-box; }

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
82% confidence
Finding

The README does not clearly warn users that supplied paper URLs may be fetched over the network and local PDF paths may be opened and processed. This missing disclosure can lead to accidental exposure of local document contents or unintended outbound requests, especially in agent environments where users may not realize that a casual prompt causes file access and remote retrieval.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

The skill directs the generated HTML to implement editing with localStorage persistence but does not clearly warn the user that edits may remain stored in the browser. This can surprise users on shared machines or managed environments where persisted content may expose notes or copied paper content to later users of the same browser profile.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
86% confidence
Finding

Several 'Best for' bullet points are written in Chinese, while the rest of the file is in English. This creates a mixed-language experience that effectively favors a specific language without explicit user opt-in or justification, which may violate language/locale policy expectations.

Content

No source excerpt is available for this finding.