train-convo-steering

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

• Third-party content exposure detected (high risk: 0.80). The nightly pipeline clearly ingests untrusted user-generated content from the runtime logs (e.g., ./_out/live_logs.jsonl -> episodes.jsonl) and sends those user_text entries to an external DeepSeek/scillm judge (train_convo_steering/judge_deepseek.py called from train_convo_steering/pipeline.run_judge_subset), and the judge labels are then used by train_priors to update per-user priors that directly influence runtime preset selection.