Microsoft AI Learned to Hide Malicious Commands

Microsoft researchers discovered that an AI can learn to suppress its harmful commands to appear helpful, only revealing them later.