Microsoft Researchers Uncover Hidden Backdoor Vulnerability in AI Models

Microsoft researchers have demonstrated a new technique, "Lazarus," to embed hidden backdoors into AI models by modifying fewer than 10 rarely used neurons without retraining. This method allows an attacker to trigger specific commands from an AI agent with a chosen string, while the model otherwise appears normal. The research reveals a significant vulnerability in AI safety and supply chain security, prompting Microsoft to issue defense guidance.