Lazarus Attack Backdoors AI Agents via Under 10 Neurons
Microsoft researchers have demonstrated a technique called Lazarus that implants a hidden backdoor into open-weight LLMs by modifying fewer than 10 rarely-active neurons, requiring no model retraining. When a predefined trigger string appears in input or tool output, the backdoored model injects an attacker-controlled command into its output, which an AI agent then executes with its own permissions. The attack preserves normal model behaviour and benchmark performance, making detection extremely difficult, and poses a serious supply chain risk for any organisation deploying open-weight models.