AI Security Maturity Model·AWS v1.0.0

Drag to explore · pinch to zoom

Enhanced Network SecurityDefense-in-depth layer: Infrastructure Capability 3 · Retrieval-augmented generation (RAG)Use case: AI that connects 👤 CustomerAWS Shared Responsibility Model: who implements and operates this control. Customer = you own it; AWS = AWS operates it and you inherit it; Shared = both have obligations.

c3.6 Keep the vector store on private connectivity

Reach the OpenSearch Serverless vector store over PrivateLink so retrieval traffic stays private and only approved services can query it.

Foundational
Enhanced
Advanced
Effort: Medium Impact: Medium GenAI scopesFrom the AWS Generative AI Security Scoping Matrix.: 3 · Pre-trained models; 4 · Fine-tuned models

Implementation guidance

How to implement on AWS

Use AWS PrivateLink to create a private connection between your VPC and the OpenSearch Serverless endpoint, and use network policy rules to specify Amazon Bedrock access. Keep the collection off public networks.

The vector store holds your data — keep it private

Retrieval reads from the vector store (commonly Amazon OpenSearch Serverless). That store contains the embeddings — and often the text chunks — of your source documents. If its endpoint is reachable from the public internet, the retrieval layer becomes an exposed copy of your knowledge base. OpenSearch Serverless controls this with network policies: rules that declare whether a collection's data-plane endpoint is public or private.

Make the collection private and bind it to your VPCEs

A network policy with AccessType: private restricts the collection to specified OpenSearch Serverless–managed VPC endpoints (and to AWS services granted private access):

[
  {
    "Description": "Private access to the RAG collection",
    "Rules": [
      {
        "ResourceType": "collection",
        "Resource": ["collection/my-rag-collection"]
      }
    ],
    "AllowFromPublic": false,
    "SourceVPCEs": ["vpce-0abc123456789def0"]
  }
]

With AllowFromPublic: false, only traffic arriving through the listed VPC endpoint(s) reaches the collection's data plane.

Let Bedrock reach it privately

A Bedrock knowledge base needs to query the collection. OpenSearch Serverless supports AWS service private access, so you can authorize the Bedrock service to reach a private collection without exposing it publicly — configured on the network policy alongside your VPCEs. This keeps the whole retrieval path (Bedrock → collection) off the public internet.

Layer the other policy types

Network policy is one of three OpenSearch Serverless policy types — pair it with:

  • an encryption policy using your CMK (see the KB encryption control), and
  • a data access policy granting only the knowledge base service role read/write on the index (see the KB service-role control).

Network keeps it private; encryption protects it at rest; data access decides who can read/write.

Naming note: replace the collection name and VPC endpoint id. Create the VPC endpoint in the subnets your workloads use, and verify AllowFromPublic is false for any collection holding sensitive embeddings.

Acceptance criteria

The vector store is reached via PrivateLink; network policies restrict callers to Bedrock/approved services; no public endpoint remains.

Practical implementation

How this control is implemented in each reference architecture:

The OpenSearch Serverless collection in the Generative AI account is reached only over PrivateLink, with network policies limiting callers to Bedrock and approved services.

  1. Create an AWS PrivateLink connection between the Generative AI account VPC and the OpenSearch Serverless endpoint.
    Generative AI account AWS PrivateLink OpenSearch Service Amazon VPC
  2. Set OpenSearch Serverless network policy rules to allow access only from Amazon Bedrock and approved services, and keep the collection off public networks.
    Generative AI account OpenSearch Service IAM
  3. Restrict the VPC endpoint and security groups to the retrieval path only.
    Generative AI account AWS PrivateLink Amazon VPC

How to validate: The vector store is reached via PrivateLink; network policies limit callers to Bedrock/approved services; no public endpoint remains.

Avoid: Leaving the collection on a public network policy for convenience. Retrieval traffic and embeddings must stay on private connectivity.

Generative AI OU (multi-account) →

The vector store is reached over PrivateLink within the account, with network policies restricting callers to the approved retrieval path.

  1. Create a PrivateLink connection between the account VPC and the OpenSearch Serverless endpoint.
    Generative AI tier (IAM role) AWS PrivateLink OpenSearch Service Amazon VPC
  2. Set network policy rules to allow only Bedrock/approved services and keep the collection off public networks.
    Generative AI tier (IAM role) OpenSearch Service IAM
  3. Scope the endpoint and security groups to the retrieval path.
    Generative AI tier (IAM role) AWS PrivateLink Amazon VPC

How to validate: The vector store is reached via PrivateLink; network policies restrict callers; no public endpoint remains.

Avoid: Exposing the collection publicly because it is 'just embeddings'. Embeddings can leak source content; keep the store private.

Single account (entry point) →

Threats addressed

  • Vector store reachable from unintended networks.
  • Retrieval traffic exposed on the public internet.

AWS services

AWS PrivateLink Amazon OpenSearch Serverless Amazon VPC

Pricing

Network and access policies add no charge. You pay for the OpenSearch Serverless collection (OCUs) and, if you create interface VPC endpoints, the per-AZ-hour plus data-processing cost. There is no extra charge for making a collection private. See Network access for Amazon OpenSearch Serverless.

Framework mappings

OWASP GenAI/LLM Top 10 (2026)

  • LLM02 Sensitive Information Disclosure

NIST AI Risk Management Framework (AI 100-1)

  • MANAGE.1 AI risks are prioritized, responded to and managed

AWS Well-Architected Generative AI Lens (Security)

  • GENSEC01-BP02 Implement private network communication between foundation models and applications

Learning resources

References

  • AWS SRA for AI — Capability 3, Amazon OpenSearch Serverless