Curated References
Links I Find Useful
A working bibliography of frameworks, whitepapers, and field reports I keep coming back to when thinking about AI security, agent architecture, and identity for AI systems.
Cloud Security Alliance — AI Research Publications
CSA's index of whitepapers, frameworks, and best-practice guides covering AI risk, governance, and secure deployment — vendor-neutral and updated regularly.
OWASP GenAI Security Project
Home of the OWASP Top 10 for LLM Applications and broader generative-AI threat models. Practical, vendor-neutral guidance for builders and reviewers.
OWASP AI Exchange
Open reference for AI threats, controls, and mitigation strategies. Cross-mapped to ISO/IEC, NIST AI RMF, and the EU AI Act for compliance alignment.
MITRE ATLAS — Adversarial ML Threat Matrix
Knowledge base of real-world attacks against AI systems, modeled in the ATT&CK framework. Tactics, techniques, and documented case studies.
CIS — Insights & White Papers
Center for Internet Security's library of practical hardening guidance, benchmarks, and emerging-tech security briefs — including AI-specific advisories.
NIST — Control Overlays for Securing AI Systems
NIST CSRC announcement of an SP 800-53 control overlay tailored to AI workloads — maps existing federal controls to AI-specific risks and lifecycle stages.
Google Cloud — Securing AI (Practitioner Resources)
Google's Secure AI Framework (SAIF) hub. Design principles, practitioner guides, and reference architectures for securing AI workloads in production.
AWS Security Blog — Data Authorization for GenAI Apps
How to apply fine-grained authorization to the data feeding GenAI apps. Covers RAG patterns, identity context propagation, and least-privilege at retrieval time.
AWS Whitepaper — Navigating the Security Landscape of GenAI
AWS reference paper covering threat models, shared-responsibility boundaries, and control selection across the GenAI lifecycle from data prep to inference.
Anthropic — Building Effective AI Agents
Anthropic's architecture-pattern guide for production agents: tool use, planning, memory, and the trade-offs between deterministic workflows and autonomous loops.
The Lethal Trifecta — Simon Willison
The canonical framing of why any agent that combines private data, exposure to untrusted content, and external communication is inherently exfiltration-prone.