Ep 16: From POCs to Production: Building a Real GenAI Security Program

In this episode of Access Granted, Nauman sits down with Ken Huang—co-author of the OWASP Top 10 for LLMs, contributor to NIST AI work, and co-chair of CSA’s AI Safety Group—to break down what practical GenAI security looks like.They cover:Why only a small fraction of organizations feel comfortable with their GenAI security postureThe three big risk buckets: prompt injection, MCP/tooling exposure, and goal manipulation / agent driftHow “shadow AI” emerges when there’s no dedicated GenAI security programA concrete framework stack: NIST AI RMF → Maestro threat modeling → OWASP AI VSS → CSA AICM + red teamingThe role of cloud provider frameworks (Google SAIF, AWS CAF-E AI, Azure guidance) and how to combine them with community standardsWhy traditional IAM (static SAML/OAuth scopes) doesn’t work for AI agents—and what task-scoped, intent-based, ephemeral access should look likeHow to think about identity lifecycle and governance for AI agents, and why “no 24/7 God mode” should be a non-negotiable anchor for CISOsIf you’re trying to move from GenAI science projects to production systems without sleepwalking into a breach—or letting an agent delete your production database—this conversation will help you define the guardrails, frameworks, and identity controls you actually need.