hackernews

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

• Third-party content exposure detected (high risk: 0.90). The skill explicitly fetches and ingests public, user-generated content (Hacker News via the Firebase API and Algolia, HN comments, and arbitrary linked article/GitHub URLs) as part of its normal workflow — see SKILL.md sections "When the user asks to chase an artifact" and the hn-deepdive.sh script which instructs the sub-agent to curl the ARTICLE_URL and run hn-cli.sh to fetch comments — so untrusted third-party content is read and can drive follow-up actions.

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

• Potentially malicious external URL detected (high risk: 0.90). The hn-deepdive.sh script builds and sends a runtime prompt to the sub-agent that explicitly instructs it to curl an arbitrary ARTICLE_URL (e.g., "curl -sL 'https://article-url.com' | pandoc -f html -t gfm-raw_html") and to run local commands, which causes fetched remote content to be injected into the agent's context and can directly control the agent's behavior — this is a runtime fetch of external content that the skill relies on for deep dives.