Guardrails and Safety Engineering for LLM Applications
For engineers who own the safety of an LLM product: design layered input, output and action guardrails, choose and tune safety classifiers, measure their false positives and false negatives, and operate them under real traffic.
The curriculum is being prepared and reviewed. Enrollment is not open yet.
About this course
Guardrails fail in two directions: they let harmful outputs through, or they block legitimate users until the product is useless. This course teaches how to build guardrails that are explicit about that trade-off. It begins with writing a policy taxonomy specific to your product and mapping it to OWASP and NIST guidance, then works through each layer: input screening for prompt injection and jailbreaks, policy classifiers (from fixed-taxonomy models such as Llama Guard 4 to bring-your-own-policy reasoning classifiers such as gpt-oss-safeguard), PII detection and redaction, output validation against schemas and business rules, grounding and citation checks, and constraints on what tools and agents may do.
You will compare programmable frameworks (NVIDIA NeMo Guardrails, Guardrails AI) with managed services (Amazon Bedrock Guardrails including Automated Reasoning checks, Azure AI Content Safety with Prompt Shields, now part of Microsoft Foundry, Google Cloud Model Armor) and decide where each fits. The final modules cover measurement and operations: building labeled safety eval sets, choosing thresholds from precision-recall curves, latency budgets for synchronous checks, streaming moderation, user-facing refusal and appeal design, and monitoring for drift and new attack patterns. Offensive red-teaming technique is covered in sec-310; this course is about building and running the defenses. The course ends in an automated exam; passing earns a Wirewalk Institute of AI certificate of completion and assessment.
What you will be able to do
Write a product-specific safety policy taxonomy and map it to OWASP Top 10 for LLM Applications and NIST AI 600-1 risk categories
Design a layered guardrail architecture spanning input, model, output and action checks with defined failure behavior
Select and configure safety classifiers, comparing fixed-taxonomy, policy-prompted and prompt-injection models on recall, precision and latency
Implement guardrails with NeMo Guardrails or Guardrails AI and with a managed service such as Bedrock Guardrails or Azure AI Content Safety
Constrain agent actions with allow-lists, argument validation and human approval for high-impact tools
Set classifier thresholds from labeled evaluation data and report false-positive and false-negative rates by category
Operate guardrails in production with streaming moderation, latency budgets, refusal UX, appeals and drift monitoring
Syllabus
6 modules · 19 lessons · about 18 hours
1
Module 1: Safety Requirements and Policy
Deciding precisely what the system must not do, and for whom.3 lessons
Harms, Misuse and Failure: Scoping Safety for a Specific Product
Writing a Policy Taxonomy Classifiers Can Enforce
Mapping to OWASP Top 10 for LLM Applications and NIST AI 600-1
2
Module 2: Guardrail Architecture
Where checks go, what they cost, and what happens when they fire.3 lessons
Defense in Depth: Input, Model, Output and Action Layers
Synchronous, Asynchronous and Streaming Checks: Latency Trade-offs
Fail-Open vs Fail-Closed and Designing the Refusal Experience
3
Module 3: Classifiers and Detectors
The models that make the decisions, and how they differ.4 lessons
Moderation Models: OpenAI Moderation and Llama Guard 4
Policy-as-Prompt Classifiers with gpt-oss-safeguard
Prompt Injection and Jailbreak Detection with Prompt Guard 2 and Prompt Shields
PII Detection and Redaction with Microsoft Presidio and Managed Services
4
Module 4: Guardrail Frameworks and Managed Services
Programmable rails versus cloud guardrail products.3 lessons
NVIDIA NeMo Guardrails: Colang Flows and Rail Types
Guardrails AI: Validators, Re-Asking and Structured Output Enforcement
Amazon Bedrock Guardrails, Azure AI Content Safety and Google Cloud Model Armor Compared
5
Module 5: Grounding and Action Guardrails
Keeping answers tied to sources and agents inside their authority.3 lessons
Grounding and Citation Checks Against Retrieved Sources
Formal and Rule-Based Checks: Automated Reasoning and Business Rules
Action Guardrails: Tool Allow-Lists, Argument Validation and Human Approval
6
Module 6: Measuring and Operating Guardrails
Proving guardrails work and keeping them working.3 lessons
Building Labeled Safety Eval Sets and Over-Refusal Tests
Threshold Selection from Precision-Recall Curves per Category
Production Monitoring: Block Rates, Appeals, Drift and New Attack Patterns
Exam and certificate
The exam unlocks once you have completed every lesson. It draws 25 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 60 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.
Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.
The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by OpenAI, Meta, NVIDIA, Amazon Web Services, Microsoft, Google, or Guardrails AI. Product and company names are trademarks of their respective owners.
A hands-on course on building no-code assistants with the builders already in your workplace tools: ChatGPT projects, custom GPTs and their successor plugins, Claude Projects and Skills, Gemini Gems, and Microsoft 365 Copilot Agent Builder and Copilot Studio. You leave with a tested, documented assistant and a plan to share it safely.
A hands-on, code-first course on the three major LLM APIs side by side: messages, streaming, structured outputs, tool calling, vision, cost, caching, batch, retries, abstraction layers, cloud routes and testing. You finish by building and hardening a small multi-vendor service.
A practitioner's course on retrieval-augmented generation: when to use it, how to ingest, chunk, embed and index documents, how to combine BM25 and vector search with reranking, and how to measure, debug and operate the result. Every module includes working Python sketches.