Four security principles for agentic AI systems

This article outlines four foundational security principles for agentic AI systems, responding to NIST's request for industry guidance on securing autonomous AI agents.

• Secure development lifecycle practices must cover both traditional software and AI components
• Traditional security controls remain fully applicable to agentic systems
• Deterministic, infrastructure-level controls external to agents are essential for security
• Agent autonomy should expand progressively based on demonstrated performance and evaluation
• AWS implements these principles through compute isolation, identity management, tool access policies, and observability
• Agents should start with human approval for high-consequence operations, earning autonomy through evidence

In summary, securing agentic AI requires extending existing security frameworks with agent-specific considerations, emphasizing external deterministic controls over prompt-based guardrails, and progressively expanding autonomy through systematic evaluation.