build-connector

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

• Third-party content exposure detected (high risk: 0.85). Phase 0 uses WebFetch to retrieve public Fivetran documentation pages at runtime (outsider-authored web content), which the workflow then scans for connector names and may incorporate into the agent’s LLM context.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly directs runtime WebFetch of Fivetran documentation pages (e.g. https://fivetran.com/docs/connectors/databases, https://fivetran.com/docs/connectors/events, https://fivetran.com/docs/connectors/files, https://fivetran.com/docs/connectors/functions, https://fivetran.com/docs/connectors/applications and slug-probe variants) and uses the fetched page content to determine connector-matching logic and next actions, so external content directly controls the agent's decision flow.