crew-dispatch

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

• Potentially malicious external URL detected (high risk: 0.90). The helper invokes runtime HTTP endpoints (e.g. "$T3_ORIGIN/api/orchestration/dispatch") to POST a thread.turn.start user message that directly injects a prompt into the orchestration thread, so the script uses that URL at runtime to control agent prompts.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.90). The prompt explicitly mandates running crewmates in "full-permission, no-sandbox" mode (bypassing normal tool/sandbox restrictions) and permits destructive host-side actions like deleting sibling clones, which encourages bypassing security boundaries and modifying host state.