Skip to content

AWS Bedrock Implementation Patterns

Purpose: Platform-specific guidance for implementing the infrastructure controls on AWS using Amazon Bedrock as the model hosting platform.
Status: Reference patterns - adapt to your specific architecture and AWS account structure.

Architecture Mapping

Framework Zone AWS Implementation
Zone 1 - Ingress Amazon API Gateway + AWS WAF + Amazon CloudFront
Zone 2 - Runtime Amazon Bedrock (invoke endpoint), Bedrock Guardrails, Amazon OpenSearch Serverless (vector read)
Zone 3 - Evaluation Separate Bedrock model invocation (Judge) in isolated account or VPC
Zone 4 - Ingestion AWS Lambda / AWS Step Functions + Amazon Bedrock Knowledge Bases (ingestion), OpenSearch Serverless (vector write)
Zone 5 - Control Plane AWS Systems Manager Parameter Store / Secrets Manager, AWS Config, IAM Identity Center
Zone 6 - Logging Amazon CloudWatch Logs, Amazon S3 (log archive), Amazon Security Lake, SIEM integration

Identity & Access (IAM Controls)

IAM-01/02: Authentication and Least Privilege

  • Use IAM Identity Center for human identities with federated SSO.
  • Use IAM Roles for all service identities - never static access keys.
  • Bedrock model invocation via IAM role-based access with resource-based policies.
  • Resource policies on Bedrock model endpoints restrict which IAM principals can invoke.
  • Separate IAM roles for: application (invoke model), guardrail service (evaluate), ingestion (manage knowledge base), admin (configure).

IAM-03: Control/Data Plane Separation

  • Separate AWS accounts for control plane (model configuration, guardrail rules) and runtime (model invocation).
  • Use AWS Organizations SCPs to prevent runtime accounts from modifying Bedrock model configurations.
  • Control plane changes require IAM Identity Center with MFA, not programmatic access.
  • Bedrock model invocation permissions ≠ Bedrock model management permissions.

IAM-04/05: Agent Tool Constraints

  • Bedrock Agents define action groups - each action group is a declared tool with specific API operations.
  • Use Lambda function resource policies to restrict which Bedrock Agent can invoke which function.
  • Implement approval routing via Step Functions for high-impact actions - the Lambda returns a "pending approval" response and waits for human input.

IAM-06: Session-Scoped Credentials

  • Use AWS STS AssumeRole with DurationSeconds set to minimum needed (15 min to 1 hour).
  • Set session tags on STS tokens to bind to agent session ID.
  • IAM condition keys (aws:PrincipalTag/SessionId) restrict token use to originating session.

Logging & Observability (LOG Controls)

LOG-01: Model I/O Logging

  • Enable Bedrock Model Invocation Logging to CloudWatch Logs and/or S3.
  • Log configuration captures: input tokens, output tokens, model ID, invocation latency.
  • For full prompt/response capture, use the S3 delivery option (CloudWatch has size limits).
  • Important: Bedrock invocation logging captures prompts and responses - ensure PII handling (LOG-09) is applied downstream.

LOG-02/03: Guardrail and Judge Logging

  • Bedrock Guardrails log trace data when enabled - includes filter decisions, blocked content, and applied policies.
  • Use the trace field in Bedrock API responses to capture guardrail evaluation details.
  • Judge invocations logged as separate Bedrock model calls with distinct logging configuration.

LOG-04: Agent Decision Chains

  • Bedrock Agent traces capture the full reasoning chain: rationale, action, observation, final response.
  • Enable trace logging on agent invocations (enableTrace: true).
  • Store traces in S3 with lifecycle policies matching LOG-08 retention requirements.

LOG-05/06: Drift and Injection Detection

  • Use Amazon CloudWatch Metrics for Bedrock invocation metrics (latency, error rates, token consumption).
  • Set CloudWatch Alarms with anomaly detection for baseline deviation (>2σ).
  • Implement custom CloudWatch Metric Filters on invocation logs for prompt injection pattern detection.
  • Feed to Amazon Security Lake for cross-source correlation.

LOG-10: SIEM Integration

  • Use Amazon Security Lake as the normalisation layer (OCSF format).
  • Export to enterprise SIEM via Security Lake subscriber or S3 export.
  • CloudWatch Logs subscription filters for real-time streaming to SIEM.

Network & Segmentation (NET Controls)

NET-01: Network Zones

  • Use separate VPCs for runtime (Zone 2) and ingestion (Zone 4).
  • Bedrock Guardrails run in AWS-managed infrastructure - use VPC endpoints (com.amazonaws.region.bedrock-runtime) for private connectivity.
  • VPC endpoint policies restrict which IAM principals can use the endpoint.
  • Use AWS PrivateLink for all Bedrock API access - no internet-routed traffic.

NET-02: Guardrail Bypass Prevention

  • Bedrock Guardrails are applied at the API level - configure guardrail ID on model invocations.
  • Use IAM policies to deny bedrock:InvokeModel without guardrail specification.
  • SCP at org level: deny Bedrock invocations that don't include a guardrailIdentifier.

NET-03: Judge Isolation

  • Judge model invoked from a separate VPC or separate AWS account.
  • Cross-account IAM role for Judge invocation with minimal permissions.
  • Judge VPC has no inbound routes from runtime VPC - data pushed via SQS or EventBridge.

NET-04: Agent Egress Controls

  • Agent Lambda functions run in VPCs with no internet gateway.
  • Outbound access via VPC endpoints for AWS services and NAT Gateway with security group restrictions for external tools.
  • Use AWS Network Firewall for domain-based egress filtering on external tool calls.
  • VPC Flow Logs capture all network traffic for NET-08 monitoring.

Data Protection (DAT Controls)

DAT-03: PII Detection

  • Amazon Comprehend for NER-based PII detection on model I/O.
  • Bedrock Guardrails PII filters for built-in detection of common PII types.
  • Custom regex patterns via Lambda@Edge or pre-processing Lambda for domain-specific PII.

DAT-04: Access-Controlled RAG

  • Bedrock Knowledge Bases support metadata filtering on retrieval.
  • Implement document-level access control by adding classification metadata at ingestion.
  • Use OpenSearch Serverless fine-grained access control for field-level security on embeddings.
  • Pre-filter approach: add metadata filter to Knowledge Base query matching user's permissions.

DAT-05: Encryption

  • Bedrock data encrypted at rest with AWS KMS - use CMK (customer-managed key) for Tier 3+.
  • OpenSearch Serverless collections encrypted with KMS.
  • S3 log buckets encrypted with separate KMS key.
  • All Bedrock API calls over TLS 1.2 via PrivateLink endpoints.

Secrets & Credentials (SEC Controls)

SEC-01/03: Vault and Context Isolation

  • Use AWS Secrets Manager for all AI system credentials.
  • Agent Lambda functions retrieve credentials from Secrets Manager at invocation - credentials not stored in environment variables.
  • Bedrock Agent action groups use Lambda execution roles - no credentials in agent prompts.

SEC-04: Credential Scanning

  • Amazon Macie for scanning S3 logs for credential patterns.
  • CodeGuru Security for scanning AI application code.
  • Custom guardrail rules in Bedrock for credential pattern detection in I/O.

Incident Response (IR Controls)

IR-04: Rollback

  • Bedrock model versions pinned - rollback by changing the model ID in configuration.
  • Bedrock Guardrail versions - create new version, point to previous known-good version.
  • Knowledge Base rollback: re-sync from known-good S3 source data.
  • AWS CodePipeline with manual approval gates for all AI component deployments.

AWS-Specific Considerations

Consideration Guidance
Bedrock Guardrails limitations Bedrock Guardrails run in AWS-managed infrastructure - you control configuration but not the execution environment. Supplement with custom guardrail logic for domain-specific requirements.
Multi-region Bedrock model availability varies by region. Design for single-region deployment per AI system instance.
Cost management Enable Bedrock model invocation metrics for cost tracking. Use SESS-01 token limits to prevent denial-of-wallet.
Model access Bedrock requires explicit model access grants per account. Use this as an additional control point (SUP-01).
CloudTrail Enable CloudTrail for all Bedrock management API calls. This covers control plane auditing (IAM-08).