New research from Microsoft Research demonstrates that reprogramming just a few dormant neurons i...

Microsoft researchers have discovered a technique called Lazarus, which allows attackers to backdoor an AI agent by weight-reprogramming fewer than 10 dormant neurons without requiring retraining. The compromised model passes standard safety evaluations and behaves normally until a specific trigger appears, at which point it can execute unauthorized commands. Microsoft is developing defenses such as AI-powered weight scanning, cryptographic model signing, and provenance verification to mitigate these risks. Open article →