Cybersecurity

Microsoft Research Reveals AI Supply Chain Vulnerabilities and How Coding Agents Are Compromised

Two recent studies by the Microsoft Research team and the AI Red Team reveal how poisoned datasets and models can be exploited to plant backdoors in coding agents and execute malicious code remotely.

2026-11-12
3 min read
135 views
certi.news Editorial Team
Microsoft Research Reveals AI Supply Chain Vulnerabilities and How Coding Agents Are Compromised

Two new studies conducted by Microsoft Defender Research and the Microsoft AI Red Team have highlighted advanced security risks threatening AI supply chains, specifically those related to coding agents and intelligent terminals. The studies demonstrated that relying solely on model behavior during standard evaluation processes does not guarantee operational safety, given the ability of attackers to plant latent malicious behaviors that trigger only when specific conditions are met.

Modern coding agents rely on broad permissions that allow them to inspect code repositories, read files, call APIs, and execute commands, which multiplies the severity of any covert compromise. The research examined two primary paths for supply chain attacks post-training: the first relies on dataset poisoning, while the second relies on direct model poisoning.

In the dataset poisoning path, an attacker plants carefully crafted examples within supervised fine-tuning data to teach the model insecure, conditional behavior. When receiving a seemingly innocent request from a user, the model dynamically searches for any available execution tool and misuses it. Tests on a custom model showed that this method achieved a success rate of approximately 78% in identical test environments, with limited ability to transfer to other environments.

The model poisoning path assumes the attacker has control over the training process before the open weights of the models are distributed. By combining fine-tuning and reinforcement learning, researchers were able to improve the reliability of backdoors while maintaining the model's overall performance. Attack success rates in unfamiliar repositories ranged between 84.1% and 94.8%, with high rates recorded when deploying models across different runtime environments such as Claude Code, Pi, and Codex CLI tools.

What changes in practice?

These findings confirm that traditional evaluations of model capabilities and post-training are insufficient to ensure a system is free of conditional backdoors. Operators face a real risk that a compromise could originate from data consumed in the victim's training pipelines or from checkpoints prepared by third parties prior to distribution.

Recommendations and preventive guidelines

To enhance protection against these threats, researchers recommended implementing strict data governance, auditing training repositories, and verifying their integrity using cryptographic hashing. The guidelines also emphasized the necessity of treating third-party checkpoints as untrusted dependencies, minimizing the tool permissions granted to agents, and isolating execution environments via virtual containers while monitoring tool activity and detecting any unexpected deviations in responses.

News source
Microsoft Security Blog
Open original source ↗
c
Author

certi.news Editorial Team

In the same category

You may also like

View all news