I am excited to announce the availability of AWS Key Management Service (AWS KMS) External Key Store. Customers who have a regulatory need to store and use their encryption keys on premises or outside of the AWS Cloud can now do so. This new capability allows you to store AWS KMS customer managed keys on…
I am excited to announce the availability of AWS Key Management Service (AWS KMS)External Key Store. Customers who have a regulatory need to store and use their encryption keys on premises or outside of the AWS Cloud can now do so. This new capability allows you to store AWS KMS customer managed keys on a hardware security module (HSM) that you operate on premises or at any location of your choice.
At a high level, AWS KMS forwards API calls to securely communicate with your HSM. Your key material never leaves your HSM. This solution allows you to encrypt data with external keys for the vast majority of AWS services that support AWS KMS customer managed keys, such as Amazon EBS, AWS Lambda, Amazon S3, Amazon DynamoDB, and over 100 more services. There is no change required to your existing AWS services’ configuration parameters or code.
This helps you unblock use cases for a small portion of regulated workloads where encryption keys should be stored and used outside of an AWS data center. But this is a major change in the way you operate cloud-based infrastructure and a significant shift in the shared responsibility model. We expect only a small percentage of our customers to enable this capability. The additional operational burden and greater risks to availability, performance, and low latency operations on protected data will exceed—for most cases—the perceived security benefits from AWS KMS External Key Store.
Let me dive into the details.
A Brief Recap on Key Management and Encryption When an AWS service is configured to encrypt data at rest, the service requests a unique encryption key from AWS KMS. We call this the data encryption key. To protect data encryption keys, the service also requests that AWS KMS encrypts that key with a specific KMS customer managed key, also known as a root key. Once encrypted, data keys can be safely stored alongside the data they protect. This pattern is called envelope encryption. Imagine an envelope that contains both the encrypted data and the encrypted key that was used to encrypt these data.
But how do we protect the root key? Protecting the root key is essential as it allows the decryption of all data keys it encrypted.
The root key material is securely generated and stored in a hardware security module, a piece of hardware designed to store secrets. It is tamper-resistant and designed so that the key material never leaves the secured hardware in plain text. AWS KMS uses HSMs that are certified under the NIST 140-2 Cryptographic Module certification program.
You can choose to create root keys tied to data classification, or create unique root keys to protect different AWS services, or by project tag, or associated to each data owner, and each root key is unique to each AWS Region.
The XKS Proxy Solution When configuring AWS KMS External Key Store (XKS), you are replacing the KMS key hierarchy with a new, external root of trust. The root keys are now all generated and stored inside an HSM you provide and operate. When AWS KMS needs to encrypt or decrypt a data key, it forwards the request to your vendor-specific HSM.
All AWS KMS interactions with the external HSM are mediated by an external key store proxy (XKS proxy), a proxy that you provide, and you manage. The proxy translates generic AWS KMS requests into a format that the vendor-specific HSMs can understand.
The HSMs that XKS communicates with are not located in AWS data centers.
To provide customers with a broad range of external key manager options, AWS KMS developed the XKS specification with feedback from several HSM, key management, and integration service providers, including Atos, Entrust, Fortanix, HashiCorp, Salesforce, Thales, and T-Systems. For information about availability, pricing, and how to use XKS with solutions from these vendors, consult the vendor directly.
In addition, we will provide a reference implementation of an XKS proxy that can be used with SoftHSM or any HSM that supports a PKCS #11 interface. This reference implementation XKS proxy can be run as a container, is built in Rust, and will be available via GitHub in the coming weeks.
Once you have completed the setup of your XKS proxy and HSM, you can create a corresponding external key store resource in KMS. You create keys in your HSM and map these keys to the external key store resource in KMS. Then you can use these keys with AWS services that support customer keys or your own applications to encrypt your data.
Each request from AWS KMS to the XKS proxy includes meta-data such as the AWS principal that called the KMS API and the KMS key ARN. This allows you to create an additional layer of authorization controls at the XKS proxy level, beyond those already provided by IAM policies in your AWS accounts.
The XKS proxy is effectively a kill switch you control. When you turn off the XKS proxy, all new encrypt and decrypt operations using XKS keys will cease to function. AWS services that have already provisioned a data key into memory for one of your resources will continue to work until either you deactivate the resource or the service key cache expires. For example, Amazon S3 caches data keys for a few minutes when bucket keys are enabled.
The Shift in Shared Responsibility Under standard cloud operating procedures, AWS is responsible for maintaining the cloud infrastructure in operational condition. This includes, but is not limited to, patching the systems, monitoring the network, designing systems for high availability, and more.
When you elect to use XKS, there is a fundamental shift in the shared responsibility model. Under this model, you are responsible for maintaining the XKS proxy and your HSM in operational condition. Not only do they have to be secured and highly available, but also sized to sustain the expected number of AWS KMS requests. This applies to all components involved: the physical facilities, the power supplies, the cooling system, the network, the server, the operating system, and more.
Depending on your workload, AWS KMS operations may be critical to operating services that require encryption for your data at rest in the cloud. Typical services relying on AWS KMS for normal operation include Amazon Elastic Block Store (Amazon EBS), Lambda, Amazon S3, Amazon RDS, DynamoDB, and more. In other words, it means that when the part of the infrastructure under your responsibility is not available or has high latencies (typically over 250 ms), AWS KMS will not be able to operate, cascading the failure to requests that you make to other AWS services. You will not be able to start an EC2 instance, invoke a Lambda function, store or retrieve objects from S3, connect to your RDS or DynamoDB databases, or any other service that relies on AWS KMS XKS keys stored in the infrastructure you manage.
As one of the product managers involved in XKS told me while preparing this blog post, “you are running your own tunnel to oxygen through a very fragile path.”
We recommend only using this capability if you have a regulatory or compliance need that requires you to maintain your encryption keys outside of an AWS data center. Only enable XKS for the root keys that support your most critical workloads. Not all your data classification categories will require external storage of root keys. Keep the data set protected by XKS to the minimum to meet your regulatory requirements, and continue to use AWS KMS customer managed keys—fully under your control—for the rest.
Some customers for which external key storage is not a compliance requirement have also asked for this feature in the past, but they all ended up accepting one of the existing AWS KMS options for cloud-based key storage and usage once they realized that the perceived security benefits of an XKS-like solution didn’t outweigh the operational cost.
What Changes and What Stays the Same? I tried to summarize the changes for you.
The way to protect access and monitor access from the AWS side is unchanged. XKS uses the same IAM policies and the same key policies. API calls are logged in AWS CloudTrail, and AWS CloudWatch has the usage metrics.
The pricing is the same as other AWS KMS keys and API operations.
XKS does not support asymmetric or HMAC keys managed in the HSM you provide.
You now own the concerns of availability, durability, performance, and latency boundaries of your encryption key operations.
You can implement another layer of authorization, auditing, and monitoring at XKS proxy level. XKS resides in your network.
While the KMS price stays the same, your expenses are likely to go up substantially to procure an HSM and maintain your side of the XKS-related infrastructure in operational condition.
An Open Specification For those strictly regulated workloads, we are developing XKS as an open interoperability specification. Not only have we collaborated with the major vendors I mentioned already, but we also opened a GitHub repository with the following materials:
The XKS proxy API specification. This describes the format of the generic requests KMS sends to an XKS proxy and the responses it expects. Any HSM vendor can use the specification to create an XKS proxy for their HSM.
A reference implementation of an XKS proxy that implements the specification. This code can be adapted by HSM vendors to create a proxy for their HSM.
An XKS proxy test client that can be used to check if an XKS proxy complies with the requirements of the XKS proxy API specification.
Other vendors, such as SalesForce, announced their own XKS solution allowing their customers to choose their own key management solution and plug it into their solution of choice, including SalesForce.
Pricing and Availability External Key Store is provided at no additional cost on top of AWS KMS. AWS KMS charges $1 per root key per month, no matter where the key material is stored, on KMS, on CloudHSM, or on your own on-premises HSM.
Architecture to AWS CloudFormation code using Anthropic’s Claude 3 on Amazon Bedrock
In this post, we explore some ways you can use Anthropic’s Claude 3 Sonnet’s vision capabilities to accelerate the process of moving from architecture to the prototype stage of a solution. Source
In this post, we explore some ways you can use Anthropic’s Claude 3 Sonnet’s vision capabilities to accelerate the process of moving from architecture to the prototype stage of a solution.
Control data access to Amazon S3 from Amazon SageMaker Studio with Amazon S3 Access Grants
In this post, we demonstrate how to simplify data access to Amazon S3 from SageMaker Studio using S3 Access Grants, specifically for different user personas using IAM principals. Source
In this post, we demonstrate how to simplify data access to Amazon S3 from SageMaker Studio using S3 Access Grants, specifically for different user personas using IAM principals.
Scalable training platform with Amazon SageMaker HyperPod for innovation: a video generation case study
In this post, we share an ML infrastructure architecture that uses SageMaker HyperPod to support research team innovation in video generation. We will discuss the advantages and pain points addressed by SageMaker HyperPod, provide a step-by-step setup guide, and demonstrate how to run a video generation algorithm on the cluster. Source
In this post, we share an ML infrastructure architecture that uses SageMaker HyperPod to support research team innovation in video generation. We will discuss the advantages and pain points addressed by SageMaker HyperPod, provide a step-by-step setup guide, and demonstrate how to run a video generation algorithm on the cluster.