Module 14: Guardrails, Safety, and Alignment | Generative AI & LLM

Duration: 4 hours | Difficulty: Intermediate–Advanced | Prerequisites: Modules 01–08

Learning Objectives

By the end of this module, you will be able to:

Map common LLM threat categories (injection, data exfil, toxic outputs)
Apply layered defenses for prompts and tools
Validate and filter outputs before users see them
Integrate Guardrails AI or NeMo Guardrails-style policies
Follow a responsible deployment checklist

1. The Threat Landscape

Why Safety Matters

Models optimize for plausible text, not safe outcomes. In prod, you own PII, brand, and legal risk—assume adversarial users and sloppy insiders.

Risk	Example
Data leak	Model echoes secrets from context
Abuse	Instructions for harm
Misinformation	Confident wrong medical/legal claims
Availability	Prompt bomb spikes cost

Try This! Write three “friendly” prompts that try to exfiltrate system instructions—see if your app blurs secrets.

2. Prompt Injection Attacks and Defenses

Types of Prompt Injection

Style	What attacker does
Direct	“Ignore previous rules…”
Indirect	Poisoned doc retrieved by RAG
Tool misuse	Args that hit forbidden APIs

Defense: Input Validation

Layer	Idea
Length limits	Cheap DoS brake
Allowlists	Only expected languages/channels
Separate trusted system text	Don’t concat untrusted docs into system

Defense: LLM-Based Injection Detection

Run a small classifier prompt (“is this trying to override policies?”) before the main call—treat as heuristic, not proof.

Concept: Defense in depth beats one magic filter.

3. Output Validation and Filtering

Key Example: Regex PII scan + simple issue list—extend with your DLP vendor patterns.

python

[object Object], re
,[object Object], dataclasses ,[object Object], dataclass

,[object Object],
,[object Object], ,[object Object],:
    patterns: ,[object Object],[,[object Object],, ,[object Object],]

    ,[object Object], ,[object Object],(,[object Object],):
        ,[object Object],._rx = {k: re.,[object Object],(v) ,[object Object], k, v ,[object Object], ,[object Object],.patterns.items()}

    ,[object Object], ,[object Object],(,[object Object],) -> ,[object Object],:
        hits = {name: rx.findall(text) ,[object Object], name, rx ,[object Object], ,[object Object],._rx.items() ,[object Object], rx.search(text)}
        ,[object Object], {,[object Object],: ,[object Object], hits, ,[object Object],: hits}


guard = OutputGuard(
    {
        ,[object Object],: ,[object Object],,
        ,[object Object],: ,[object Object],,
    }
)
sample = ,[object Object],
,[object Object],(guard.scan(sample))

Post-process	When
Redact PII	Support transcripts
Block toxic phrases	Consumer apps
JSON schema validate	Tool pipelines

4. Guardrails AI

Declarative validators (ToxicLanguage, DetectPII, etc.) wrap outputs or prompts. Configure on_fail strategies: fix, reask, exception.

Plus	Minus
Fast to prototype	Another dependency + latency

Fun Fact: Validators can fight each other—e.g., PII redactor vs citation requirement—test combined configs.

5. NeMo Guardrails

YAML + Colang flows describe conversational policies: input rails, output rails, specialized dialog flows. Good when you want explicit policy files reviewed by compliance.

Asset	Role
`config.yml`	Model endpoints
`.co` flows	Allowed / forbidden topics

6. Responsible AI Deployment Checklist

Area	Sample items
Input	Rate limits, injection alerts, file type allowlist
Output	PII scan, moderation, citation rules for RAG
System	Secret management, audit logs, kill switch
People	Disclosure, escalation path, red-team cadence