event-prospecting-pipeline

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

• Third-party content exposure detected (high risk: 0.85). Step 2 uses the luma-event-attendees sub-skill to “scrape the public attendee list from the event URL”; that runtime-fetched attendee text is outsider-authored free-form content from a public web page and is ingested into the agent/LLM context for filtering/enrichment.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill calls the runtime endpoint "$PEPPER_CLOUD_URL/api/state" (via state_read/state_write/state_append) to fetch the required strategy/icp.md content, and that fetched content is used at runtime to filter, score, and seed prompts to downstream skills (i.e., it directly controls agent behavior and is a required dependency).