hallmark

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

• Third-party content exposure detected (high risk: 0.85). In hallmark study URL mode, the skill performs runtime WebFetch of a user-supplied URL and parses the returned HTML/CSS to extract design facts, which can include outsider-authored free text (e.g., page body/copy) that is then ingested into the agent’s LLM context for diagnosis/build.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill's "hallmark study" verb explicitly WebFetches user-supplied http:// or https:// URLs at runtime and parses their HTML/CSS as input to produce a diagnosis and (optionally) to "build with this DNA", meaning remote page content can directly influence the agent's prompts/output (user-supplied http(s) URLs fetched via WebFetch).