Menu

Microsoft researchers backdoored an AI agent by editing fewer than 10 neurons - no retraining needed

Reading 0:00
15s threshold

Microsoft Research published a study this week that should worry anyone deploying open-weight models: a technique they call "Lazarus" can plant a working backdoor in an AI agent by reprogramming fewer than 10 dormant neurons in the underlying LLM - no retraining, no poisoned data, just a small edit to the saved weights.

How it works: every large model has neurons that almost never fire. Lazarus repurposes a handful of them as trigger detectors. The edited model behaves normally and passes standard safety evaluations - until the trigger appears. The trigger can be something as mundane as a date: in one demo, once the clock passed midnight into 2027, the agent, in the middle of an ordinary task, emitted a command that downloaded and ran attacker code.

The numbers, from Microsoft's Nov 10 post: - Nemotron-Terminal-8B: 7 neurons out of 12,000+ in the final feed-forward layer, trigger worked 98.8% of the time - Qwen3.6-27B: 6 neurons, success 51-99.5% depending on setup - Gemma-4-31B-IT was also tested

Two…

Continue reading — create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.