Microsoft Discovers Lazarus Dormant Neuron Backdoor Technique

On November 11, 2026, Microsoft Research released first-hand research revealing the discovery of a backdoor technique named Lazarus. The technique requires no retraining of the model—by reprogramming fewer than 10 dormant neurons in a large language model, an AI Agent built on that model can be made to execute attacker commands when hidden trigger conditions arise. This finding shows that the underlying models AI Agents rely on may harbor a very small number of highly concealed neuron-level backdoors, posing a security risk of being exploited to manipulate Agent behavior. Current reports only cover the discovery and technical description of the technique; there have been no follow-up reports on specific defense solutions proposed by Microsoft, responses from other organizations, or actual attack cases. The situation remains under continuous observation.

AI overview of reporting · Updated 1 h前

Not enough continuous observations to show a trend yet.