Microsoft researchers show dormant neurons can backdoor AI agents
Microsoft Research says it found "Lazarus," a backdoor technique that reprograms dormant neurons in an AI model so an otherwise normal agent runs malicious commands when a hidden trigger appears. The researchers write that reprogramming fewer than 10 dormant neurons is enough to backdoor an agent, requires no training, and leaves the model passing standard evaluations. The post says the attack needs only write access to a checkpoint, so internally developed models are equally exposed, and it says Microsoft Foundry verifies model provenance before onboarding open-weight models.