meta-ad-policy-checker

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

• Third-party content exposure detected (medium risk: 0.30). In SKILL.md Phase 3 (and Phase 1/2), the workflow ingests outsider-authored ad copy inputs (headline/body/CTA and optional visual description/targeting context) and reasons over them against fetched Meta policy text, so user-submitted free text is directly consumed at runtime.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly fetches Meta's ad-standards index at runtime (https://transparency.meta.com/policies/ad-standards/) and injects the fetched policy text as ground-truth citations used to generate verdicts, so the external content directly controls the agent's outputs.