model-poisoning-reviewer

What problem does it solve?

This Skill helps you assess whether training data, feedback loops, or ingested content can be poisoned to corrupt a model's future behavior, introduce backdoors, or steer answers through tainted ground truth.

Core Features & Use Cases

• Training and fine-tuning review: Checks dataset provenance, curation, label integrity, and validation gates before model updates.
• Feedback-loop abuse detection: Reviews whether open ratings, corrections, or conversational feedback could be mass-signaled to bias the next training cycle.
• Ingestion integrity analysis: Evaluates whether external documents indexed into a knowledge base can be used to manipulate retrieval outcomes or embed malicious triggers.
• Use case: A team retrains weekly on customer feedback and public web content; this Skill identifies poisoning paths and recommends controls like provenance tracking, anomaly detection, holdout evaluation, and purge/rollback readiness.

Quick Start

Review our fine-tuning pipeline and feedback loop for poisoning risks, trace any attacker-reachable input to corrupted behavior, and list the controls that would prevent it.