hallmark

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

• Third-party content exposure detected (high risk: 0.85). Outsider free text can enter the LLM context via hallmark study <URL> URL-mode: the skill performs runtime WebFetch of a user-supplied external page and parses its HTML/CSS into readable text for extraction/diagnosis, which is outsider-authored content.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill's "study" verb explicitly WebFetches user-supplied http:// or https:// URLs at runtime (e.g., any https://example.com passed to hallmark study) and parses the returned HTML/CSS as input to its diagnosis/build, so remote pages fetched at runtime directly influence the agent's prompts/output.