icp-research

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

• Third-party content exposure detected (high risk: 0.90). The skill's workflow explicitly instructs automated collection and ingestion of public, user-generated content (e.g., agents/voc-collector-agent.md requires collecting 15+ quotes from Reddit, G2, Twitter, LinkedIn and SKILL.md's Pre-Dispatch "Auto-scan first" step reads marketing/pricing pages), and those third‑party inputs are used by downstream agents (synthesis/critic/routing) to drive decisions and re-dispatches—creating a clear channel for untrusted content to influence agent behavior.