Security Considerations for Artificial Intelligence Agents (Perplexity / NIST CAISI Response)

arXiv 2603.12230·high signal

Perplexity's formal response to NIST's 2025-0035 RFI documents security observations from operating production agentic systems at millions of users / thousands of enterprises. Covers how agent architectures break code-data separation and authority boundaries, requiring new security primitives beyond user-input filtering. Serves as an authoritative industry reference for organizations architecting agent security policy today.