New AI Malware PoC Reliably Evades Microsoft Defender

A security researcher demonstrated a lightweight, reinforcement-learning–trained model that autonomously generates malware designed to evade Microsoft Defender for Endpoint; after iterative rewards tied to Defender alerts the model achieved roughly an 8% success rate for producing fully evasive malware, can run on consumer GPUs, and was developed for a modest cost, raising concerns that criminals could adopt similar methods to scale EDR evasion.