twg-bench-lite

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

• Third-party content exposure detected (medium risk: 0.30). In scripts/benchmark-lite/runner.mjs, the required runtime workflow directly injects user-supplied --prompt/--prompt-file text into both live arms as part of the sub-agent input (child.stdin.end(armInstructions(label, prompt, routePlan))), so outsider-authored free text is ingested by the agent without any prerequisite selection of a specific item.