model-poisoning-reviewer

Detect data poisoning by reviewing dataset provenance, curation gates, and label integrity to identify corrupted behavior paths before model updates. This process checks training sets for malicious inputs and validates ground truth to prevent backdoor insertion.

RLHF feedback abuse occurs when open ratings, corrections, or conversational feedback are mass-signaled to bias the next training cycle. Analyzing feedback loops identifies these poisoning paths and recommends controls like anomaly detection and identity limits.

External documents indexed for RAG can be manipulated to alter retrieval outcomes or embed malicious trigger phrases. Ingestion integrity analysis evaluates whether indexed content can poison knowledge bases and recommends purge and rollback controls.

Review backdoor trigger-phrase risks by tracing attacker-reachable inputs through ingestion and feedback loops to corrupted model behavior. This analysis requires severity-ranked findings with concrete poisoning paths and validation holdout evaluation.

Controls to prevent data poisoning include provenance tracking, anomaly detection, holdout evaluation, and purge or rollback readiness. These controls secure training sets, fine-tuning data, and feedback loops against behavior-shaping manipulation.