hallmark

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

• Third-party content exposure detected (high risk: 0.75). Yes—hallmark study <screenshot | URL> in URL mode fetches arbitrary public web page HTML/CSS via WebFetch at runtime and then uses that extracted page content to generate the diagnosis/build, which can include outsider-authored free text (e.g., visible copy or embedded strings) that the LLM ingests.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill's study verb explicitly WebFetches user-supplied URLs (http:// and https://) at runtime and parses the returned HTML/CSS into the model context to extract "DNA", so external pages fetched during runtime can directly influence the agent's prompts and outputs.