Beyond the Prompt: Infrastructure Hardening for Scalable Agentic AI

While the security community remains hyper-focused on LLM-specific threats like prompt injection, real-world enterprise deployments reveal that the most critical vulnerabilities often lie in the underlying architecture. This session explores the hard-learned lessons from several Generative AI (GenAI) deployments that I have worked on, moving past theoretical risks to actually addressing the infrastructure failures that break production environments. We will deconstruct three primary architectural pillars: The Compute Layer: How under-provisioned containerized backends create self-inflicted Denial of Service (DoS) and how to implement auto-scaling guardrails. The Networking Layer: Navigating the "VPC Lockout" phenomenon where aggressive data exfiltration controls break Retrieval-Augmented Generation (RAG) workflows. The IAM Layer: Solving the "God-Mode" dilemma in Agentic AI by implementing automated least-privilege for service accounts with broad agency. Attendees will leave with a technical blueprint for securing RAG and Agentic workflows, including specific strategies for Layer 7 filtering, VPC Service Control tuning, and Infrastructure-as-Code (IaC) patterns to remediate over-privileged AI service accounts. Key Takeaways for Attendees Root-Cause Methodology: Techniques for using cloud-native monitoring to distinguish between model hallucinations and infrastructure-induced latency. Network Policy Tuning: How to design "allow-lists" for AI agents and external vector databases without compromising data exfiltration boundaries. Agentic IAM Governance: Practical application of IAM recommenders and IaC to strip "God-Mode" permissions from AI service accounts. Defense-in-Depth: Implementing Layer 7 filtering to block malicious inputs before they reach the model endpoint.