skills
coreyhaines31/marketingskills/skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 1f837447b20b
- Files read
- 460
This project is a collection of guides and helper scripts that help AI coding agents handle marketing tasks like SEO audits, paid ads, conversion tracking, social media, and competitor research. The biggest risk is that a target website can trick its competitor research script into probing your local network, and several workflows tell the AI to read untrusted web pages or follow live setup instructions from outside websites while having access to your terminal and browser.
What it can reach
- Files
- Reads and writes marketing notes, ad creative assets, research reports, and Google Tag Manager files in your project folder (such as .agents/) and your home folder (~/marketing-plans/).
- Network
- Fetches public websites, sitemaps, DNS and certificate logs, app store reviews, social feeds, and public APIs (like GitHub, npm, Reddit, Wayback Machine, and SEC), and queries any analytics, ad, or CRM tools you connect.
- Shell commands
- Runs bundled Python and Node scripts (build_recipe.py and teardown.mjs) and instructs the agent to run terminal commands like git, curl, unpinned npx packages, media tools like ffmpeg, and Playwright.
- Secrets
- Reads optional API keys and tokens from environment variables (including GITHUB_TOKEN, SCRAPECREATORS_API_KEY, AHREFS_API_KEY, and ELEVENLABS_API_KEY) and guides connecting credentials for ad, analytics, and CRM services.
- Browser
- Uses browser automation tools like Playwright to inspect pages, capture screenshots, verify tracking tags, record local app demos, read logged-in LinkedIn or X feeds, and fill out directory submission forms for your approval.
Findings
-
MediumServer-side request forgery via untrusted sitemaps, GrowthBook hosts, and redirects
skills/reverse-engineer-marketing/scripts/teardown.mjs:533
When teardown.mjs analyzes a target domain, sitemapUrls() extracts Sitemap URLs from the target's robots.txt and <loc> URLs from <sitemapindex> XML and fetches them directly without checking the URL scheme, hostname, or resolved IP address. In addition, cmdExperiments() fetches any https:// host matched from apiHost in page HTML, and all fetch() calls follow HTTP redirects automatically. A malicious or compromised target website can use these paths to make the user's machine send HTTP GET requests to localhost ports, private LAN services, or cloud metadata endpoints.
Fix: Validate every URL fetched from robots.txt, sitemap indexes, page HTML, and redirect hops so that only https: URLs on public IP addresses are allowed, and restrict sitemap fetches to the target domain or explicitly trusted hosts.
-
MediumSkill instructs agent to fetch and follow live external install instructions
skills/conversion-tracking/references/server-side.md:163
Although conversion-tracking/SKILL.md states that fetched pages are data rather than instructions, references/server-side.md instructs the agent to fetch https://www.tracklution.com/agent-install.md live at runtime and follow its instructions directly to write code into the user's project. If that remote URL or domain is compromised or altered, an attacker can inject arbitrary instructions and code changes into the user's workspace.
Fix: Replace the live remote instruction directive with a reviewed, versioned summary stored in the local skill directory, and treat any fetched vendor documentation strictly as untrusted data for human review.
-
LowSkills instruct downloading and running unpinned npx packages at runtime
Several skills instruct running unpinned third-party npm packages directly via npx, including npx is-agentic and npx afdocs check during AI SEO audits (skills/ai-seo/SKILL.md and references/agent-readiness.md) as well as npx hyperframes and npx create-video@latest (skills/video/SKILL.md and skills/ad-creative/references/generative-tools.md). Executing unpinned npx packages downloads and runs the latest published code from npm on the user's machine, exposing local files and environment secrets if an upstream package is compromised.
Fix: Pin exact, audited package versions (such as package@x.y.z), prefer passive HTML and HTTP checks over third-party CLIs during audits, and require explicit user confirmation before executing external packages.
-
LowDirectory submission workflow prioritizes third-party /llms.txt instructions and MCP endpoints
skills/directory-submissions/SKILL.md:147
While directory-submissions/SKILL.md warns to treat directory pages as untrusted data and check DNS resolution, it also tells the agent to prefer /llms.txt instructions or MCP endpoints advertised by each directory across a catalog of roughly 200 sites, including small indie directories that can lapse or change ownership while still resolving in DNS. Connecting to untrusted MCP endpoints or following procedural instructions in third-party /llms.txt files exposes the agent to prompt injection.
Fix: Do not connect to MCP endpoints or follow procedural instructions from untrusted third-party directories' /llms.txt files; treat remote docs solely as untrusted data and restrict outbound submissions to user-approved web forms or explicitly trusted registries.
-
LowProtocol-relative script URL in Microsoft UET GTM recipe
skills/conversion-tracking/assets/gtm-recipes/send/microsoft.json:31
Unlike the Meta, LinkedIn, and TikTok GTM recipes—which load their tracking scripts explicitly over https://—the Microsoft Ads UET recipe loads //bat.bing.com/bat.js using a protocol-relative URL. If a site imports this GTM recipe and serves any page over plain HTTP or previews a local HTML file, the browser can request the third-party script without TLS instead of enforcing HTTPS.
Fix: Replace '//bat.bing.com/bat.js' with 'https://bat.bing.com/bat.js' so the tag always loads the UET script over HTTPS.
