Back Issues/Search Home → Calendar → Archive → RSS → Subscribe → Services → Vendors → Blog → Current Issue → Popular → Newsletters (LinkedIn, Substack)

All issues › Volume 344, Issue 2 › IT Vendor News › Microsoft

Trained to Betray: How Post-Training Supply Chain Attacks Hijack Coding Agents

Microsoft, Thursday, November 12th, 2026

Microsoft research shows poisoned fine-tuning data and poisoned model checkpoints can backdoor coding agents to run malicious commands.

Microsoft Defender Research and the Microsoft AI Red Team present two studies of post-training supply chain attacks on coding agents.

In dataset poisoning, a small number of crafted fine-tuning examples teach a model to find any command-execution tool and misuse it on a routine user request, reaching about 78% attack success on in-distribution harnesses and transferring to an unseen one.

In model poisoning, an adversary who controls training before releasing an open-weight checkpoint combines supervised fine-tuning with reinforcement learning, reaching 84.1% to 94.8% average success on unseen repositories.

Triggers can look benign, so prompt-injection defenses may miss them; the authors call for controls over data pipelines, model adoption and tools.

more →  ·  More from Microsoft →