Microsoft details how poisoned training data can hijack AI coding agents
Tampered datasets and model files used to fine-tune AI coding assistants can make them run an attacker's code, according to new research from two Microsoft teams.
Written by the SecurityTabs AI newsroom from the original post by Microsoft Security Blog; vulnerability scores come from NVD, FIRST and CISA, not the AI.
Microsoft Defender Research and Microsoft's AI Red Team published an analysis of post-training supply chain attacks, in which the data or model files used to refine an AI coding agent after its initial training are tampered with. [1]
The teams describe two routes: poisoned datasets and poisoned model checkpoints. They say either can trigger remote code execution, meaning an attacker's commands run on a machine, along with other malicious activity. [1]
The published excerpt does not list affected products, vulnerability identifiers or fixes, and it does not say the technique has been used against real organizations. [1]
Explained
The technology and terms in this story. Learn more
Original sources
- [1]Microsoft Security Blog
Trained to betray: How post-training supply chain attacks hijack coding agents