Microsoft Researchers Reveal 'Lazarus' Backdoor That Hides Malicious Commands in Dormant AI Neurons

At a glance

Microsoft researchers found a backdoor technique called Lazarus that reprograms a few dormant neurons in an AI model. The model behaves normally until a hidden trigger appears, then runs malicious commands. The risk is highest for AI agents that have access to tools or terminals.

Key points

1. 1

   Changing fewer than 10 dormant neurons is enough to plant the backdoor, and no retraining is needed.

2. 2

   The model passes standard safety tests, so the backdoor is hard to detect.

3. 3

   A trigger can be as simple as a date change.

4. 4

   Once triggered, the agent downloads and runs attacker code with the access it was given.

Why it matters

As AI agents gain access to files and systems, checking where a model came from and confirming it has not been altered is becoming a core security task.

Also reported by

• AI 매터스MS 리서치, 백도어 기법 ‘라자루스’ 공개…뉴런 10개 미만 조작