Building AI Defenses

What you will be able to do

• Implement layered guardrails with input and output filtering that block prompt injection, unsafe content, and data leakage before they reach a user or a downstream tool
• Build an input and output filtering pipeline from classifier-based and rule-based controls (moderation, PII redaction, structured-output validation) and measure its false-positive and false-negative rates
• Harden a model against adversarial and jailbreak attacks using the NIST AI 100-2 taxonomy and MITRE ATLAS, and run a red-team evaluation to prove the hardening works
• Secure a retrieval-augmented generation pipeline against data poisoning, indirect prompt injection, and cross-tenant leakage with document-level access control
• Protect model API keys, service credentials, and model artifacts with a secrets manager, short-lived credentials, key rotation, and egress controls
• Design an AI gateway that enforces authentication, rate and budget limits, and least-privilege tool access on every model call
• Map real attack techniques to specific engineering controls and communicate the residual risk of an AI system to stakeholders
• Assemble the individual controls into a defense-in-depth reference architecture you can apply to a new AI product

What is inside

4 modules, 8 lessons. Each module ends in a graded quiz and most carry an assignment.

1. 01

   Guardrails and Input/Output Filtering

   Build a layered guardrail system around the model boundary. On the way in you detect prompt injection, moderate content, and redact PII. On the way out you validate structure, moderate responses, and enforce a guardrail policy. You wire classifier-based and rule-based controls together, decide fail-open versus fail-closed, and measure with precision and recall whether the controls actually work.

   2 lessons · 5 quiz questions

2. 02

   Hardening Models for Adversarial Robustness

   Filters can be bypassed, so the model itself must resist attack. This module teaches adversarial machine learning the way an AI security engineer needs it. You will structure jailbreaks, adversarial examples, and prompt injection with the NIST AI 100-2 taxonomy and MITRE ATLAS, apply the hardening that lowers their success rate (adversarial training, system-prompt defense, and defense in depth), and prove it holds with an automated red-team evaluation.

   2 lessons · 5 quiz questions · assignment

3. 03

   Securing Retrieval-Augmented Generation

   Retrieval-augmented generation is the most common production AI architecture, and it opens a new attack surface: the knowledge base and the retrieval path. This module secures RAG end to end. You will see how poisoned documents and indirect prompt injection hijack the model through retrieved content, then build the defenses that stop them: document-level access control, retrieval-time authorization, tenant isolation, content sanitization, and provenance, so answers stay trustworthy and scoped to the caller.

   2 lessons · 5 quiz questions

4. 04

   Protecting Keys, Secrets, and the Inference Stack

   An AI system is only as safe as the credentials and artifacts beneath it. This module secures the substrate: the model provider API keys and service credentials your stack depends on, and the model weights and prompts that are themselves sensitive assets. You will vault secrets with rotation and short-lived credentials, harden model artifacts against theft and tampering, front every model call with an AI gateway that enforces authentication, egress control, and rate and budget limits, and assemble all four modules of the course into a single defense-in-depth reference architecture.

   2 lessons · 5 quiz questions · assignment