New Q1 2027 certification windows are open — lock in your audit before your next release. Schedule a meeting

Amazon Web Services

Secure AI deployment on AWS

The configuration we expect for Amazon Bedrock, SageMaker AI and the AWS services around them — and exactly how we verify it before issuing a certificate.

Quick answer

A secure AWS AI deployment gives each workload an IAM role limited to approved Amazon Bedrock models, reaches Bedrock through VPC interface endpoints (PrivateLink), enforces Bedrock Guardrails through the bedrock:GuardrailIdentifier IAM condition, filters Knowledge Base retrieval by user entitlements, requires confirmation for sensitive agent actions, and turns on model invocation logging, CloudTrail and GuardDuty across the organization.

Services in scope

  • Amazon Bedrock
  • Bedrock Guardrails
  • Bedrock Knowledge Bases
  • Bedrock Agents / AgentCore
  • Amazon SageMaker AI
  • Amazon OpenSearch Serverless
  • AWS Lambda (agent tools)
  • Amazon S3 (AI data)
  • AWS KMS
  • CloudTrail, GuardDuty, Security Hub

Legend: Required must pass for certification. Assessed is tested and reported with recommendations, and becomes required where your data classification demands it.

Identity & access

ControlRecommended configurationHow we verify
No long-lived credentials
Required
Workloads use IAM roles (Lambda, ECS task, EC2 instance profile, EKS Pod Identity). People sign in through IAM Identity Center. Long-term IAM user access keys and long-term Bedrock API keys are not used for AI workloads.Credential report and IAM Access Analyzer review; search for access keys used against Bedrock in CloudTrail.
Least-privilege model access
Required
bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream scoped to approved model and inference profile ARNs. No bedrock:* for application roles.Policy analysis of every principal with Bedrock permissions.
Organization guardrails
Assessed
Service control policies restrict Bedrock to approved regions and models, and deny deleting the invocation logging configuration.SCP review in AWS Organizations.
Secrets
Required
Third-party AI keys are stored in AWS Secrets Manager with rotation — never in Lambda environment variables, code or notebooks.Configuration review of Lambda, ECS and SageMaker resources.

Network isolation

ControlRecommended configurationHow we verify
PrivateLink endpoints
Required
VPC interface endpoints for bedrock-runtime and bedrock-agent-runtime with private DNS, and endpoint policies limited to your principals and actions.VPC endpoint and endpoint policy review.
Endpoint-only access
Assessed
IAM policies deny Bedrock calls that don't arrive through your endpoints, using aws:SourceVpce.Policy analysis.
SageMaker isolation
Assessed
Endpoints and training jobs run in your VPC with network isolation. Notebook instances have direct internet access and root access disabled.SageMaker resource configuration review.

Data protection

ControlRecommended configurationHow we verify
AI data buckets
Required
S3 buckets holding training, fine-tuning, knowledge base or log data have Block Public Access on, SSE-KMS encryption, and bucket policies limited to specific roles.Bucket configuration and policy review.
Data residency
Required
Regions and cross-region inference profiles are chosen deliberately to match where data may be processed.Inference profile and region usage review (CloudTrail).
Customer managed keys
Assessed
AWS KMS customer managed keys for custom models, knowledge bases, agents and AI data stores. Required for regulated data.KMS key and key policy review.
Sensitive data discovery
Assessed
Amazon Macie scans buckets that feed RAG and fine-tuning.Macie job and findings review.

Guardrails

ControlRecommended configurationHow we verify
Bedrock Guardrails
Required
Guardrails configured with the prompt attack filter, content filters, denied topics and sensitive information filters (mask or block PII). Contextual grounding checks on RAG applications.Guardrail configuration review and adversarial testing of live applications.
Guardrails enforced by IAM
Required
Application roles may only invoke models with the approved guardrail, using the bedrock:GuardrailIdentifier condition key.Policy analysis; test invocation without a guardrail is denied.
Untrusted input tagging
Assessed
Retrieved documents and user-supplied content are wrapped in guardrail input tags so they're evaluated for prompt attacks.Code review and indirect injection testing.

Rate limits & cost

ControlRecommended configurationHow we verify
Per-consumer limits
Required
Per-user or per-tenant throttling at the API tier (Amazon API Gateway usage plans, AWS WAF rate-based rules or an AI gateway).Configuration review and controlled load testing.
Quotas & budgets
Assessed
Bedrock service quotas monitored; AWS Budgets alerts on AI spend.Quota and budget review.

Knowledge Bases (RAG)

ControlRecommended configurationHow we verify
Entitlement filtering
Required
Knowledge base retrieval applies metadata filters derived from the authenticated user's entitlements on the server — never values supplied by the model.Code review; cross-user and cross-tenant retrieval testing.
Separated knowledge
Required
Public, internal and per-tenant content live in separate knowledge bases or are strictly partitioned.Knowledge base and data source inventory.
Vector store access
Assessed
OpenSearch Serverless collections have network and data access policies restricted to the knowledge base role.Collection policy review.

Agents & tools

ControlRecommended configurationHow we verify
Least-privilege actions
Required
Action group Lambda functions have narrow execution roles. Resource-based policies allow invocation only by the specific agent.Lambda role and resource policy review.
User confirmation
Required
User confirmation is enabled for sensitive actions, and the backend independently authorizes every action against the real user.Agent configuration review; agent abuse testing.

Logging & threat detection

ControlRecommended configurationHow we verify
Invocation logging
Required
Bedrock model invocation logging enabled in every region in use, to an encrypted, access-restricted CloudWatch Logs group or S3 bucket.Logging configuration check per account and region.
CloudTrail
Required
An organization trail covering all regions, with log file validation, delivered to a protected central account.Trail configuration review.
Threat detection
Required
Amazon GuardDuty enabled organization-wide. AWS Security Hub aggregates findings and routes them to your responders.Service status and finding-routing review.
Config rules
Assessed
AWS Config rules or conformance packs track drift on AI-related resources.Config compliance review.

Reference configuration

An application role policy that only allows approved models — and only with your guardrail attached:

IAM policy
{
  "Version": "2012-10-17",
  "Statement": [{
    "Sid": "InvokeApprovedModelsWithGuardrail",
    "Effect": "Allow",
    "Action": ["bedrock:InvokeModel", "bedrock:InvokeModelWithResponseStream"],
    "Resource": "arn:aws:bedrock:us-east-1::foundation-model/APPROVED-MODEL-ID",
    "Condition": {
      "StringEquals": {
        "bedrock:GuardrailIdentifier": "arn:aws:bedrock:us-east-1:111122223333:guardrail/GUARDRAIL-ID:1"
      }
    }
  }]
}

A service control policy that keeps Bedrock in approved regions and protects invocation logging. If you use cross-region inference profiles, include every region in the profile's geography.

SCP
{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "DenyBedrockOutsideApprovedRegions",
      "Effect": "Deny",
      "Action": "bedrock:*",
      "Resource": "*",
      "Condition": { "StringNotEquals": { "aws:RequestedRegion": ["us-east-1", "us-west-2"] } }
    },
    {
      "Sid": "ProtectInvocationLogging",
      "Effect": "Deny",
      "Action": "bedrock:DeleteModelInvocationLoggingConfiguration",
      "Resource": "*"
    }
  ]
}

Check yourself: run these in each account and region you use:

AWS CLI
# Is model invocation logging on? (empty output = off)
aws bedrock get-model-invocation-logging-configuration

# Which guardrails exist?
aws bedrock list-guardrails

# Is Bedrock reachable privately?
aws ec2 describe-vpc-endpoints \
  --filters Name=service-name,Values=com.amazonaws.us-east-1.bedrock-runtime

How we certify an AWS deployment

  1. Read-only accessYou deploy a cross-account IAM role in each in-scope account. It requires an external ID and has only the AWS managed SecurityAudit and ViewOnlyAccess policies.
  2. Full AI inventoryDiscovery of Bedrock, SageMaker AI, knowledge bases, agents, guardrails and AI data buckets across every in-scope account and region — including ones nobody listed.
  3. Configuration auditEvery resource and policy is checked against this baseline, cross-referenced with Security Hub and IAM Access Analyzer findings.
  4. Attack testingAdversarial testing of the applications and agents running on the infrastructure: prompt injection, data leakage, retrieval filtering and tool abuse.
  5. Remediate, verify, certifyYou receive findings with exact policy and setting changes. We re-verify, issue your certificate and evidence pack, and re-audit every quarter.

Frequently asked questions

How do you enforce Amazon Bedrock Guardrails for every request?

Use the bedrock:GuardrailIdentifier IAM condition key in the policies attached to your application roles. Requests to invoke a model that don't specify the approved guardrail are denied, so a developer can't accidentally — or deliberately — call a model without it.

Is Amazon Bedrock model invocation logging enabled by default?

No. Model invocation logging must be turned on per account and region, sending logs to CloudWatch Logs and/or Amazon S3. CloudTrail records API activity, but not the prompt and response content. Enable invocation logging, encrypt the destination, and restrict who can read it.

Does Amazon Bedrock send our data to model providers?

According to AWS documentation, Amazon Bedrock doesn't store or log prompts and completions for its own purposes, doesn't use them to train AWS or third-party models, and doesn't share them with model providers. Your own configuration — logging destinations, knowledge base sources and cross-region inference — still determines where your data goes, which is what we verify.

What access does CertiSafe AI need to certify our AWS environment?

A cross-account IAM role in each in-scope account that trusts our audit account, requires an external ID, and has the AWS managed SecurityAudit and ViewOnlyAccess policies attached. The role is removed when the audit ends.

Certify your AWS AI deployment

Most AWS environments can be audited within two weeks of a scoping call.

Schedule a Meeting