loop-builder

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

• Third-party content exposure detected (high risk: 0.70). The required workflow includes a combo-trigger gate script that fetches work from an external source at runtime (e.g., the example fetches Intercom conversations from https://api.intercom.io/...), and the resulting free-form conversation content/fields can be passed into the agent prompt as the “batch” when run === true, creating an indirect prompt-injection path from outsider-authored messages.

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

• Potentially malicious external URL detected (high risk: 0.90). templates/gate-script-example.js performs a runtime fetch to https://api.intercom.io/conversations/search and uses the returned conversations as the batch that determines whether the agent wakes and what is passed into its prompt, so remote data directly controls agent behavior.