efficient-web-research

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

• Third-party content exposure detected (high risk: 0.85). The skill’s runtime workflow includes fetching arbitrary public web pages via read_url_content (URL Protocol / Search Protocol / Multi-URL Protocol), which can ingest outsider-authored free text from those pages into the agent’s LLM context.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly instructs runtime fetches from GitHub API (e.g., https://api.github.com/repos/{owner}/{repo}) and other arbitrary URLs via read_url_content/raw.githubusercontent, decoding and injecting fetched file/page content into the model context to drive answers, so remote content can directly control prompts.