Why AI agents are dangerous - What secure architectures do better

AI agents are seen as the next step in automation, but initial practical experience shows that as soon as LLMs are allowed to act independently, new areas of attack arise.