infra-engineer

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.90). This skill explicitly performs web scraping and browsing of public sites and arbitrary URLs (e.g., references/browser-rendering.md's "AI-Powered Web Scraper" that visits https://news.ycombinator.com and other examples that call page.goto(message.body.url)), then reads page.content() and feeds that untrusted, user/third-party content to an AI for extraction, which clearly exposes the agent to indirect prompt injection.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill includes runtime headless-browser examples that fetch external pages (e.g., page.goto('https://news.ycombinator.com')) and then pass the scraped content into an AI prompt or execute page JS via Puppeteer (page.evaluate), meaning an external URL is used at runtime to control prompts and execute remote code.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.70). This skill contains explicit privileged operations (e.g., "sudo ./aws/install"), global installs (npm -g, brew), and curl|bash installers and asks the agent to configure system credentials and install CLIs — instructions that direct the agent to perform elevated, state-changing actions on the host.