agent-observability-auto-experiment

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

• Third-party content exposure detected (medium risk: 0.30). The skill loads outsider-authored free text at runtime from Datadog LLM-Obs datasets/traces (e.g., local_dataset_path, dataset_id, ml_app, trace_ids) into data.jsonl and then feeds the untrusted datapoint input/output into the LLM judge via build_judge_prompt, creating an indirect prompt-injection surface.