Improved client-side encryption: Explicit KeyIds and key commitment
November 1, 2021: AWS KMS is replacing the term customer master key (CMK) with AWS KMS key and KMS key. The concept has not changed. To prevent breaking changes, AWS KMS is keeping some variations of this term. More info.
I’m excited to announce the launch of two new features in the AWS Encryption SDK (ESDK): local KeyId filtering and key commitment. These features each enhance security for our customers, acting as additional layers of protection for your most critical data. In this post I’ll tell you how they work. Let’s dig in.
The ESDK is a client-side encryption library designed to make it easy for you to implement client-side encryption in your application using industry standards and best practices. Since the security of your encryption is only as strong as the security of your key management, the ESDK integrates with the AWS Key Management Service (AWS KMS), though the ESDK doesn’t require you to use any particular source of keys. When using AWS KMS, the ESDK wraps data keys to one or more AWS KMS keys (KMS keys) stored in AWS KMS on encrypt, and calls AWS KMS again on decrypt to unwrap the keys.
It’s important to use only KMS keys you trust. If you encrypt to an untrusted KMS key, someone with access to the message and that KMS key could decrypt your message. It’s equally important to only use trusted KMS keys on decrypt! Decrypting with an untrusted KMS key could expose you to ciphertext substitution, where you could decrypt a message that was valid, but written by an untrusted actor. There are several controls you can use to prevent this. I recommend a belt-and-suspenders approach. (Technically, this post’s approach is more like a belt, suspenders, and an extra pair of pants.)
The first two controls aren’t new, but they’re important to consider. First, you should configure your application with an AWS Identity and Access Management (IAM) policy that only allows it to use specific KMS keys. An IAM policy allowing Decrypt on “Resource”:”*” might be appropriate for a development or testing account, but production accounts should list out KMS keys explicitly. Take a look at our best practices for IAM policies for use with AWS KMS for more detailed guidance. Using IAM policy to control access to specific KMS keys is a powerful control, because you can programmatically audit that the policy is being used across all of your accounts. To help with this, AWS Config has added new rules and AWS Security Hub added new controls to detect existing IAM policies that might allow broader use of KMS keys than you intended. We recommend that you enable Security Hub’s Foundational Security Best Practices standard in all of your accounts and regions. This standard includes a set of vetted automated security checks that can help you assess your security posture across your AWS environment. To help you when writing new policies, the IAM policy visual editor in the AWS Management Console warns you if you are about to create a new policy that would add the “Resource”:”*” condition in any policy.
The second control to consider is to make sure you’re passing the KeyId parameter to AWS KMS on Decrypt and ReEncrypt requests. KeyId is optional for symmetric KMS keys on these requests, since the ciphertext blob that the Encrypt request returns includes the KeyId as metadata embedded in the blob. That’s quite useful—it’s easier to use, and means you can’t (permanently) lose track of the KeyId without also losing the ciphertext. That’s an important concern for data that you need to access over long periods of time. Data stores that would otherwise include the ciphertext and KeyId as separate objects get re-architected over time and the mapping between the two objects might be lost. If you explicitly pass the KeyId in a decrypt operation, AWS KMS will only use that KeyId to decrypt, and you won’t be surprised by using an untrusted KMS key. As a best practice, pass KeyId whenever you know it. ESDK messages always include the KeyId; as part of this release, the ESDK will now always pass KeyId when making AWS KMS Decrypt requests.
A third control to protect you from using an unexpected KMS key is called local KeyId filtering. If you explicitly pass the KeyId of an untrusted KMS key, you would still be open to ciphertext substitution—so you need to be sure you’re only passing KeyIds that you trust. The ESDK will now filter KeyIds locally by using a list of trusted KMS keys or AWS account IDs you configure. This enforcement happens client-side, before calling AWS KMS. Let’s walk through a code sample. I’ll use Java here, but this feature is available in all of the supported languages of the ESDK.
Let’s say your app is decrypting ESDK messages read out of an Amazon Simple Queue Service (Amazon SQS) queue. Somewhere you’ll likely have a function like this:
Now, when you create a KmsMasterKeyProvider, you’ll configure it with one or more KeyIds you expect to use. I’m passing a single element here for simplicity.
Decrypt the message as normal. The ESDK will check each encrypted data key against the list of KeyIds configured at creation: in the preceeding example, the single KMS key in keyArn. The ESDK will only call AWS KMS for matching encrypted data keys; if none match, it will throw a CannotUnwrapDataKeyException.
(See our documentation for more information on how encryption context provides additional authentication features!)
We recommend that everyone using the ESDK with AWS KMS adopt local KeyId filtering. How you do this varies by language—the ESDK Developer Guide provides detailed instructions and example code.
I’m especially excited to announce the second new feature of the ESDK, key commitment, which addresses a non-obvious property of modern symmetric ciphers used in the industry (including the Advanced Encryption Standard (AES)). These ciphers have the property that decrypting a single ciphertext with two different keys could give different plaintexts! Picking a pair of keys that decrypt to two specific messages involves trying random keys until you get the message you want, making it too expensive for most messages. However, if you’re encrypting messages of a few bytes, it might be feasible. Most authenticated encryption schemes, such as AES-GCM, don’t solve for this issue. Instead, they prevent someone who doesn’t control the keys from tampering with the ciphertext. But someone who controls both keys can craft a ciphertext that will properly authenticate under each key by using AES-GCM.
All of this means that if a sender can get two parties to use different keys, those two parties could decrypt the exact same ciphertext and get different results. That could be problematic if the message reads, for example, as “sell 1000 shares” to one party, and “buy 1000 shares” to another.
The ESDK solves this problem for you with key commitment. Key commitment means that only a single data key can decrypt a given message, and that trying to use any other data key will result in a failed authentication check and a failure to decrypt. This property allows for senders and recipients of encrypted messages to know that everyone will see the same plaintext message after decryption. If you’re interested in the technical details and cryptographic analysis of the construction used for key commitment in the AWS Encryption SDK, we are using the construction detailed in Shay Gueron, Key Committing AEADs, IACR ePrint 2020.
Key commitment is on by default in version 2.0 of the ESDK. This is a breaking change from earlier versions. Existing customers should follow the ESDK migration guide for their language to upgrade from 1.x versions of the ESDK currently in their environment. I recommend a thoughtful and careful migration.
AWS is always looking for feedback on ways to improve our services and tools. Security-related concerns can be reported to AWS Security at email@example.com. We’re deeply grateful for security research, and we’d like to thank Thai Duong from Google’s security team for reaching out to us. I’d also like to thank my colleagues on the AWS Crypto Tools team for their collaboration, dedication, and commitment (pun intended) to continuously improving our libraries.
Want more AWS Security how-to content, news, and feature announcements? Follow us on Twitter.