Four security principles for agentic AI systems

AWS summarizes four foundational security principles for agentic AI—secure development lifecycle covering AI and traditional components; continuation of traditional security controls; enforcement via deterministic external controls (a "security box"); and progressively earned autonomy—then maps these to architectural building blocks (compute isolation, identity and access management, tool gateway and policy enforcement, observability, and model execution environment) exemplified in Amazon Bedrock AgentCore.