
Advanced GenAI & LLM Security: Guardrails + eBPF Defense
Udemy · English
About this course
- Language
- English
- Category
- Provider
- Udemy
- Data updated
AI applications fail in two places at once: in the words, and in the machine.A single prompt can talk a model into leaking data or calling a tool it shouldn't. And once a model server is compromised, no prompt filter can see what happens next inside the container: a shell spawning, model weights being copied, a connection opening to an attacker's server. Most AI security courses cover only one of these. This course teaches both, and how to join them. You will build a complete two-plane defense system, step by step, in Python and eBPF.Plane 1 — Application guardrailsMap your AI attack surface with the OWASP Top 10 for LLM Applications and MITRE ATLAS.Build input sanitizers for direct, indirect and encoded prompt-injection attacks. Add an LLM-as-a-judge intent check and output data-loss prevention that redacts secrets and personal data. Apply Llama Guard safety categories and NeMo-style conversational rails. Secure RAG pipelines against poisoned documents with tenant isolation, provenance checks and similarity anomaly guards. Catch system-prompt leakage in streamed output with canary tokens. Gate AI agent tool calls and MCP servers with a five-stage tool gate and human approval for risky actions. Plane 2 — Linux kernel runtime defenseLearn eBPF from first principles, write probes in C, and stream kernel events through ring buffers. Watch AI containers at runtime: model-weight loading, GPU driver calls, process trees and network traffic. Use LSM BPF hooks and XDP to block dangerous file writes and outbound connections inside the kernel. Train an Isolation Forest on syscall behaviour to catch zero-day anomalies without labelled data. Map kernel events to Kubernetes…











