23FIAP MLET Course — LLM Security, Guardrails & Compliance
Role: Course Author / Instructor
Executive summary
Authored an 8-class graduate course on securing LLM-based systems — threats, guardrails, data protection, red teaming, LGPD/GDPR/EU AI Act compliance, LLMOps observability, fairness and secure production deployment — anchored on a fictional banking assistant.
- NeMo Guardrails
- PyRIT
- Garak
- Microsoft Presidio
- Microsoft Agent Framework
- Azure AI Foundry
- promptfoo
FIAP's ML Engineering postgraduate program needed a dedicated discipline on LLM security and governance, since generative-AI systems introduce new threat classes (prompt injection, jailbreak, excessive agency) and regulatory obligations (LGPD, GDPR, EU AI Act) that traditional AppSec courses don't cover.
- Design 8 classes progressing from LLM threats/vulnerabilities to guardrails and policy-as-code, data protection and tool isolation, security testing/red teaming, privacy regulation, LLMOps governance/monitoring, ethics/bias/validation, and secure production deployment.
- Anchor every class on one evolving fictional system so controls are learned as concrete architectural decisions, not abstract rules.
- Document architecture, data-flow and deployment diagrams plus ADRs for every major security decision.
I built the course around a fictional bank ("Banco Aurora") and its LLM assistant ("Aurora Assist"), which evolves class by class from a vulnerable prototype into a governed, audited system. Each class pairs a specific threat or control — e.g. prompt injection, PII isolation with Microsoft Presidio, red teaming with PyRIT/Garak, a DPIA under LGPD/GDPR — with hands-on tooling, and I wrote supporting ADRs, business cases and architecture diagrams (system, data-flow, deployment) so students see how each control changes the assistant's architecture. I set up a `uv`-managed reproducible environment across all eight classes.
Published as FIAP's course material for the LLM Security discipline; the expected result is student learning through the Aurora Assist case study, not a production deployment.
Using one evolving fictional system instead of disconnected examples made abstract regulatory requirements (LGPD, EU AI Act) and security controls (guardrails, red teaming) land as concrete architectural decisions students could trace class by class.
Architecture diagram coming soon
Drop an image at assets/img/diagrams/fiap-llm-security-guardrails.png