Microsoft Researchers Reveal How Tiny Weight Edits Can Hijack AI Models

Researchers at Microsoft Corp. have demonstrated that attackers can plant a hidden backdoor inside an artificial intelligence (AI) model by altering fewer Microsoft researchers just demonstrated a surgical way to backdoor AI models without retraining them. By editing fewer than 10 rarely used neurons, a method called Lazarus, attackers can force an AI agent to execute hidden commands when a specific trigger appears, all while leaving standard benchmark scores intact. Even more notable: the core concept was discovered by an AI coding agent, highlighting how automated tools could lower the barrier for supply-chain attacks on open-weight models. #AISecurity #CyberSecurity #MachineLearning #AI #InfoSec