supply-chain
Trained to betray: How post-training supply chain attacks hijack coding agents
Medium
Summary
This report from Microsoft Research details two distinct attack vectors targeting coding and terminal agents – dataset poisoning and model poisoning – that can be deployed by attackers with limited control over the post-training pipeline. Model poisoning, on the other hand, allows attackers to embed backdoors during model training and then distribute the compromised model, achieving high success rates even on unseen repositories. The research emphasizes the importance of robust controls over training data and checkpoint evaluation to prevent these attacks.
Summary written automatically in our own words from the original article, which belongs to its publisher and remains the reference. It may contain errors. Sources & data
