Security, compliance, and governance in Amazon Bedrock — data protection, access control, and regulatory considerations.


Data Privacy Guarantees

Bedrock provides strong data isolation:

GuaranteeDescription
No training on your dataYour prompts/responses are never used to train or improve foundation models
Data stays in your accountInputs and outputs don’t leave your AWS account
No data retentionBedrock doesn’t store your prompts or responses (you control any logging)
Model isolationYour inference runs in isolated environments

Key Point: This is a major differentiator from using provider APIs directly (OpenAI, Anthropic) — your data is truly private.


Encryption

At Rest

ComponentEncryption
Fine-tuned modelsAWS KMS (customer-managed or AWS-managed keys)
Knowledge Base dataVector store encryption (OpenSearch, Aurora, etc.)
S3 dataS3 encryption with KMS

In Transit

  • All API calls use TLS 1.2+
  • VPC endpoints use private connectivity

Network Security

FeatureDescription
VPC EndpointsPrivate connectivity, no public internet
PrivateLinkKeep traffic within AWS network
Security GroupsControl inbound/outbound traffic
No public endpoint requiredCan run entirely in private subnets

Best Practice: Use VPC endpoints for production workloads to avoid public internet exposure.


Access Control (IAM)

Fine-grained access control:

Control LevelWhat You Can Restrict
Model accessWhich models a user/role can invoke
Feature accessKnowledge Bases, Agents, Guardrails separately
ActionsInvokeModel, CreateAgent, etc.
ResourcesSpecific model ARNs, specific agents

Example IAM Policies

ScenarioPolicy Approach
Dev team can only use LlamaRestrict to specific model ARNs
Prod can use all, dev limitedSeparate roles with different permissions
Block fine-tuning in prodDeny CreateModelCustomizationJob
Only specific agentsResource-based restrictions

Audit & Monitoring

ServiceWhat It Captures
CloudTrailAll API calls (who invoked what model, when)
CloudWatch LogsModel invocation logs (optional, you enable)
CloudWatch MetricsLatency, token counts, errors, throttling
ConfigTrack Bedrock resource configurations

Important Point: CloudTrail logs all Bedrock API calls for compliance and audit purposes.


Compliance Certifications

Bedrock inherits AWS compliance:

CertificationStatus
SOC 1, 2, 3✅ Compliant
ISO 27001, 27017, 27018✅ Compliant
HIPAA✅ Eligible (with BAA)
PCI DSS✅ Compliant
FedRAMP✅ In scope (varies by region)
GDPR✅ Data processing compliant
CSA STAR✅ Compliant

Note: Compliance applies to the Bedrock service. Your application built on Bedrock must also be compliant.


Data Residency

AspectHow It Works
Region selectionYou choose which AWS Region to use
Data stays in regionPrompts/responses processed in your chosen region
Cross-region considerationsSome models may only be available in specific regions

Responsible AI

Bedrock provides tools for responsible AI deployment:

ToolPurpose
GuardrailsFilter harmful content, PII, prompt attacks
Model evaluationTest for bias, toxicity, accuracy
Automated ReasoningPrevent hallucinations
Audit loggingTrack all model usage

When to Use Bedrock (Security Perspective)

ScenarioRecommendation
Regulated industry (healthcare, finance)Bedrock (HIPAA, PCI compliance)
Data must stay in specific regionBedrock (region selection)
No data sharing with model providersBedrock (data isolation guarantee)
Need full audit trailBedrock (CloudTrail integration)
Private network requiredBedrock (VPC endpoints)
Quick prototype, less sensitive dataDirect APIs may be simpler

TL;DR

  • Data never used for training — your prompts are private
  • Encryption: KMS at rest, TLS in transit
  • Network: VPC endpoints for private connectivity
  • IAM: Fine-grained control per model, action, resource
  • Audit: CloudTrail logs every API call
  • Compliance: SOC, HIPAA, PCI, ISO, GDPR, FedRAMP

Bottom line: Bedrock is designed for enterprise security requirements.


Resources

Bedrock Security
Security documentation and best practices.

AWS Compliance
Full list of AWS compliance certifications.