Microsoft studies show post-training backdoors can hijack coding agents
On 12 November 2026, Microsoft Defender Research and the Microsoft AI Red Team described two post-training attacks that can make coding agents misuse tools after a routine request. Dataset poisoning hit about 78% success on familiar harnesses; model poisoning reached 84.1% to 94.8% on unseen repositories.
- Microsoft Defender Research and the Microsoft AI Red Team published the studies on 12 November 2026.
- Dataset poisoning taught models to find and misuse any available command-execution tool, with about 78% success on in-distribution harnesses and transfer to an unseen harness.
- Model poisoning, where the adversary controls training before releasing an open-weight checkpoint, averaged 84.1% to 94.8% success on unseen repositories.