AI security lab · Supply Chain Attacks

Compromised Training Pipeline

CI/CD pipelines for ML training are complex and involve many stages. Compromising a single stage — data preprocessing, training, evaluation, or deployment — can poison the entire output model. **Challenge:** An ML team's GitHub Actions training pipeline has been compromised. A modified workflow file injects a data poisoning step between preprocessing and training. Identify the compromised stage and explain how the attack persists undetected.