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 two studies on 12 November 2026 of post-training supply-chain attacks that can plant a backdoor in a coding agent. A backdoor is concealed behavior that stays inactive until a trigger appears. In a conventional language model, that activation might only alter the written reply. In an agent, it can lead the model to call tools, run commands, change files, or reach data.
Coding and terminal agents already look through repositories, open files, call APIs, and run commands, frequently inside sensitive development environments. Once a backdoor fires, that autonomy can affect the system around the agent even as the agent still seems to be finishing the user’s harmless task. The trigger may not look hostile. It can be an ordinary user request.
Organizations are deploying more specialized AI agents, and Microsoft argues that trust cannot rest only on a model acting as expected in evaluation or normal use. In Models as Insider Risks in the Super Intelligence Era, Satya Nadella argues that trustworthy enterprise AI needs independent observability and controls that sit outside the model.
The studies cover attackers with different levels of access. Dataset poisoning slips a small number of crafted examples into a fine-tuning dataset so a victim-trained model learns an unsafe conditional behavior. The model does not merely store one fixed tool call. It learns to locate any available command-execution tool and misuse it when a routine user request appears, including a paraphrased one. That attack reached approximately 78% success on in-distribution harnesses and also transferred to an unseen harness.
Model poisoning assumes an adversary who controls training before an open-weight checkpoint is distributed. That position lets the adversary raise backdoor reliability while limiting the loss of general performance, by pairing supervised fine-tuning (SFT) with reinforcement learning (RL). Average success on unseen repositories in that setting reached 84.1% to 94.8%, and the backdoor showed further transfer across coding harnesses.
Microsoft says the findings reinforce controls over training-data pipelines, model adoption, and tools.