Domain 3: AI Safety, Security, and Governance (20%)

← Domain 2 · Next: Domain 4 →

Exam Tip

This domain is 20% of the exam. The key topics are Guardrails (know all four filter types and that they apply to BOTH inputs and outputs), IAM least-privilege, and distinguishing CloudTrail (audit/compliance) from CloudWatch (operational monitoring). These distinctions appear in nearly every governance question.

---

3.1 Guardrails for Amazon Bedrock

What Guardrails Do

Guardrails apply safety controls to both the input prompt and the output response of a foundation model. They act as a filter layer around the FM.

User Input → [Guardrail Input Filter] → FM → [Guardrail Output Filter] → Response

Guardrails do NOT automatically apply — you must pass guardrailIdentifier and guardrailVersion in every API call where you want them active.

Guardrail Filter Types

Filter Type	What It Controls	Example
Content Filters	Block/allow harmful categories at configurable severity thresholds	Hate speech, violence, sexual content, insults
Denied Topics	Define topics the FM must refuse to discuss	Competitor products, political opinions, legal advice
Word Filters	Block specific words or phrases (exact match)	Profanity lists, brand-specific blocked terms
PII Redaction	Detect and mask/block personally identifiable information	Names, emails, SSNs, credit cards, phone numbers
Contextual Grounding Check	Verify the response is grounded in the retrieved source context — blocks hallucinated answers	RAG applications where factual accuracy is required

Configuring a Guardrail

1. Create a Guardrail in the Bedrock console or via API
2. Configure filter types and sensitivity levels for each category
3. Version the Guardrail (you can maintain multiple versions)
4. Apply in your API call: pass guardrailIdentifier + guardrailVersion to InvokeModel

Exam Trap

Guardrails filter BOTH inputs AND outputs — not just one direction. A common wrong answer claims Guardrails only filter the model's response. They evaluate the user's prompt AND the model's answer.

Also: Guardrails are not applied automatically to all Bedrock calls. You must explicitly include them in each API request.

There are five filter types — Content Filters, Denied Topics, Word Filters, PII Redaction, and Contextual Grounding Check. The exam may list only four as distractors.

PII Detection Modes

Mode	Behavior
Redact	Replace PII with a placeholder (e.g., [EMAIL]) — request/response still proceeds
Block	Reject the entire request or response if PII is detected

Denied Topics vs. Other Filters

Denied topics are the right choice when the organization wants to block a specific subject area based on business policy, not just harmful language.

Examples:

• investment advice
• stock recommendations
• cryptocurrency trading
• competitor strategy discussion

Why denied topics matter:

• They let you define prohibited subjects in natural language
• They work at the topic / intent level, not just exact keyword matching
• They can return a controlled fallback such as: "I cannot provide investment advice."
• They apply consistently across supported Bedrock model interactions

Use denied topics when:

• The business wants to prohibit domain-specific conversations
• The prohibited content is not inherently hateful, violent, or unsafe, but still off-limits by policy

Do not confuse these with other filters:

• Content filters = harmful categories like hate, violence, sexual content, insults, profanity
• PII filters = privacy protection for sensitive personal data
• Word filters = exact word/phrase blocking, which can be too blunt and cause false positives

TIP

If the scenario says "block investment advice but still allow normal finance-related conversation when appropriate", the best answer is usually Denied Topics, not word filters.

Guardrails vs. Other Bedrock Features

Feature	Primary Purpose	Not the Right Answer When...
Guardrails for Amazon Bedrock	Apply safety policies such as denied topics, content filters, PII redaction, and grounding checks across model inputs/outputs	The requirement is primarily retrieval, prompt storage, or workflow orchestration
Knowledge Bases	Implement RAG by connecting FMs to internal data sources	The requirement is to block unsafe topics or enforce content policy
Prompt Management	Create, store, version, and reuse prompts	The requirement is independent safety filtering or PII redaction
Bedrock Agents	Execute multi-step workflows with tools and company systems	The requirement is granular safety policy enforcement

TIP

If the question is asking "which Bedrock feature enforces safety policies consistently across models?" the answer is Guardrails, not Knowledge Bases, Prompt Management, or Agents.

---

3.2 IAM & Access Control

Least-Privilege IAM for Bedrock

Service roles required:

• Bedrock needs an IAM service role to access S3 during Knowledge Base ingestion
• Knowledge Bases use a service-linked role to access OpenSearch Serverless

Key IAM actions for Bedrock:

bedrock:InvokeModel
bedrock:InvokeModelWithResponseStream
bedrock:InvokeAgent
bedrock:Retrieve
bedrock:CreateKnowledgeBase

Resource-Based Policies

• Cross-account access: Use resource-based policies on Bedrock resources to allow access from another AWS account
• Condition keys: Restrict access by bedrock:ModelId to enforce which models a principal can invoke

Lambda Execution Roles for Bedrock

When a Lambda function calls Bedrock, the correct pattern is to attach an IAM execution role to the function — not hardcode credentials.

How it works:

• Lambda automatically provides the execution role's temporary credentials to the function at runtime via the instance metadata service
• No credentials to manage, rotate, or accidentally expose
• Credentials are short-lived and scoped to the role's permissions

Minimum required policy for a Lambda that calls Bedrock:

{
  "Effect": "Allow",
  "Action": "bedrock:InvokeModel",
  "Resource": "arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-3-sonnet-*"
}

Scoping Resource to a specific model ARN enforces least-privilege — the function can only invoke that model, nothing else.

Best Practices

• Grant bedrock:InvokeModel scoped to specific model ARNs — not Resource: "*"
• Use IAM execution roles for Lambda (not IAM users with long-lived access keys) — temporary credentials are provided automatically
• Apply bedrock:ModelId condition keys as an additional constraint when resource-level scoping is not sufficient
• Never hardcode AWS credentials inside Lambda function code or environment variables

---

3.3 VPC Endpoints & Private Connectivity

Why Use VPC Endpoints?

Without a VPC endpoint, Bedrock API calls route over the public internet. VPC Endpoints (AWS PrivateLink) keep traffic within the AWS network backbone — no public internet routing.

Use VPC Endpoints when:

• Data residency or compliance requirements prohibit public internet routing
• Compute (Lambda, EC2, ECS) is inside a VPC
• You need to ensure FM traffic never leaves the AWS network

Bedrock VPC Endpoint Types

Endpoint	Service	Used For
bedrock	Amazon Bedrock control plane	Creating Knowledge Bases, Agents, Guardrails
bedrock-runtime	Amazon Bedrock runtime	InvokeModel, InvokeModelWithResponseStream
bedrock-agent-runtime	Amazon Bedrock Agent runtime	InvokeAgent calls

TIP

The exam may ask which endpoint is needed for inference calls. That's bedrock-runtime — not the control plane endpoint.

---

3.4 Traceability & Auditability

CloudTrail vs. CloudWatch

Service	Purpose	What It Records
AWS CloudTrail	Audit / compliance logging	Every Bedrock API call (who called what, when, from where)
Amazon CloudWatch	Operational monitoring	Metrics (latency, token counts, errors), alarms, dashboards

CloudTrail is the right answer when the question involves:

• "Audit trail for regulatory compliance"
• "Who made this API call and when?"
• "Log all Bedrock API activity for security review"

CloudWatch is the right answer when the question involves:

• "Monitor latency and throttling errors"
• "Set an alarm if token usage exceeds a threshold"
• "Operational dashboards for the GenAI application"

Model Invocation Logging

• Bedrock can log all model invocations (input prompts + output completions) to S3 or CloudWatch Logs
• Enable via: Bedrock console → Settings → Model invocation logging
• Use case: Traceability of what was sent to the FM and what it returned — required for AI governance and compliance

Exam Trap

CloudTrail vs CloudWatch — the exam frequently swaps these as distractors. CloudTrail = who called Bedrock (audit). CloudWatch = how Bedrock is performing (operations). Model Invocation Logging captures the actual content of prompts and responses — different from both.

---

3.5 Responsible AI

AWS Responsible AI Principles

AWS grounds its AI services in these principles — know them for governance questions:

Principle	What It Means
Fairness	AI systems should not produce discriminatory outputs or amplify biases
Explainability	Stakeholders should be able to understand why a model produced a given output
Privacy & Security	Customer data must be protected; models should not leak sensitive information
Safety	Models should not produce harmful content or take harmful actions
Controllability	Humans must be able to intervene, override, or shut down AI systems
Veracity & Robustness	Models should produce accurate, consistent outputs across varied inputs
Governance	Organizations need policies, processes, and accountability structures for AI use

AWS AI Service Cards

AWS publishes AI Service Cards for its managed AI services (including Rekognition, Textract, Comprehend, etc.). Each card documents:

• Intended use cases and out-of-scope uses
• Known limitations and biases identified during testing
• Performance across different demographic groups
• Recommended safeguards for responsible deployment

Exam Relevance

If a question asks "how do you understand the limitations and intended use cases of an AWS AI service," the answer is the AI Service Card for that service — not the API documentation.

Bias & Fairness

Bias sources in GenAI systems:

• Training data bias: if training data over-represents certain groups or perspectives, the model reflects that
• Prompt bias: poorly designed prompts can elicit biased outputs
• Feedback loop bias: RLHF-trained models may amplify rater preferences

Mitigation approaches:

• Evaluate model outputs across different demographic inputs before production deployment
• Use Bedrock Guardrails to block harmful or discriminatory content categories
• Use Model Evaluation to run diverse prompt sets and measure output consistency
• Apply human review for high-stakes decisions (hiring, lending, healthcare)

Human Oversight

• High-stakes decisions (legal, medical, financial) must include a human review step — do not fully automate
• Bedrock Agents can be configured with a human-in-the-loop confirmation step before executing irreversible actions
• Model Invocation Logging enables post-hoc review of all FM inputs and outputs for accountability

Exam Trap

Guardrails and safety filters are not a substitute for human oversight in high-stakes scenarios. The exam may present Guardrails as sufficient — they are a necessary layer but not the complete answer when human judgment is required.

---

3.6 Data Classification & Pre-Ingestion Security

Amazon Macie

Amazon Macie is a managed data security service that uses machine learning to automatically discover, classify, and protect sensitive data stored in Amazon S3.

What Macie does:

• Scans S3 buckets at scale and identifies files containing PII (SSNs, names, addresses, medical record numbers, credit cards)
• Generates inventory reports showing which buckets and objects contain sensitive data
• Provides risk scores and findings surfaced in a Security Hub-integrated dashboard
• Sends findings to EventBridge for automated remediation workflows

GenAI relevance — pre-ingestion audit:

Before connecting an S3 bucket as a Bedrock Knowledge Base data source, you should verify it does not contain unintended sensitive data. Macie is the right tool for this:

S3 Buckets (raw documents)
    ↓ Macie scans at scale
Findings Report (which buckets/files contain PII, risk scores)
    ↓ Security team reviews
Safe buckets approved → Bedrock Knowledge Base ingestion

Macie vs. Comprehend vs. Bedrock Guardrails PII

	Amazon Macie	Amazon Comprehend	Bedrock Guardrails PII
When	Before ingestion — audit S3 inventory	At processing time — analyze text	At inference time — inside Bedrock call
What	Discovers which S3 objects contain PII	Detects and extracts PII entities from text	Redacts or blocks PII in prompts/responses
Output	Findings report, risk scores, inventory	Structured entity annotations	Masked/blocked content in API response
Best for	"Audit which S3 buckets are safe to use as RAG sources"	Pre/post-processing text in a pipeline	Real-time PII control within Bedrock

Exam Scenario

"Before building a RAG knowledge base from company S3 buckets, how do you identify which buckets contain sensitive PII?" → Amazon Macie

Macie operates at the S3 bucket/object level before data ever reaches Bedrock. Comprehend and Guardrails operate on text content during or after ingestion.

---

← Domain 2 · Next: Domain 4 →