Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons

Summary

Microsoft researchers showed that an AI model can be backdoored by editing fewer than 10 rarely used neurons, without retraining. When a chosen trigger appears, the model outputs an attacker's command, which an agent using the model would run with the agent's own permissions, so the damage depends on what the agent is allowed to do. The attacker must be able to change the model's weights, and the researchers call the method Lazarus.

Topics

• AI agents

Related items

• InfoTop GenAI security resources — November 2026Same vendor · Adversa AI Blog
• MediumNovember 2026 Patch Tuesday: 63 Critical Microsoft fixes, plus an Outlook zero day that allows theft of hashed credentialsSame vendor · CSO Online
• MediumDetecting sleeper agents hiding in AI model neuronsSame vendor · Microsoft Security Blog
• MediumCVE-2026-83953 GitHub Copilot for Jetbrains Information Disclosure VulnerabilitySame vendor · MSRC Security Update Guide
• InfoCVE-2026-119990 Azure OpenAI Elevation of Privilege VulnerabilitySame vendor · MSRC Security Update Guide