Updated 30th March 2017 to reflect updated information (see comments), add additional links and add some clarifying text about why misuse-resistance is useful.
With the impending release of the ForgeRock Identity Platform, I thought I’d spend some time writing up a few of the bits of OpenAM 14 that I was directly involved with creating. One of my last acts before leaving FR to go solo, was to put in place the first phase of modernising AM’s aging system credential encryption scheme. Before I start, I should say that this encryption scheme is not used for encrypting user passwords (which are hashed by the LDAP user store, not AM). Instead, this scheme is used for encrypting various system credentials (passwords for SMTP servers, HMAC shared secrets, etc) in the config store and in exported system configurations and in a few other places.
The original (and still default) encryption method was first mentioned in Dante’s Inferno. Actually it dates from the original iPlanet codebase from the mid-90s, and uses correspondingly ancient cryptographic algorithms (MD5 and DES). It is best to regard it as providing only limited obfuscation of credentials, rather than any true security guarantees, and the advice has always been to secure the config store by traditional means (TLS, access controls) rather than rely on this encryption. Still, we can do much better than this now, so AM 14 ships with a new AESWrapEncryption scheme that provides significantly improved security:
Certainly, there are lots of potential gotchas in the specs, and it is easy for somebody without experience to shoot themselves in the foot using these standards. I agree with pretty much all of the criticisms levelled against the standards. They are too complicated with too many potentially insecure options. It is far too easy to select insecure combinations or misconfigure them. Indeed, much of the advice in my earlier article can be boiled down to limiting which options you use, understanding what security properties those options do and do not provide, and completely ignoring some of the more troublesome aspects of the spec. If you followed my advice of using “headless” JWTs and direct authenticated encryption with a symmetric key, you’d end up not far off from the advice of just encrypting a JSON object with libsodium or using Fernet.
So in that sense, I am already advocating for not really using the specs as-is, at least not without significant work to understand them and how they fit with your requirements. But there are some cases where using JWTs still makes sense:
If you need to implement a standard that mandates their use, such as OpenID Connect. In this case you do not have much of a choice.
If you need to interoperate with third-party software that is already using JWTs. Again, in this case you also do not have a choice.
You have complex requirements mandating particular algorithms/parameters (e.g. NIST/FIPS-approved algorithms) and don’t want to hand-roll a message format or are required to use something with a “standard”. In this case, JWT/JOSE is not a terrible choice, so long as you know what you are doing (and I hope you do if you are in this position).
If you do have a choice, then you should think hard about whether you need the complexity of JWTs or can use a simpler approach that takes care of most of the choices for you or store state on the server and use opaque cookies. In addition to the options mentioned in the referenced posts, I would also like to mention Macaroons, which can be a good alternative for some authorization token use-cases and the existing libraries tend to build on solid foundations (libsodium/NaCl).
So, should you use JWT/JOSE at all? In many cases the answer is no, and you should use a less error-prone alternative. If you do need to use them, then make sure you know what you are doing.
Update (20th April, 2017): I’ve noticed that this article gets by far the most daily hits on my blog. This worries me that people are using this code as a template for building real ECDHE key agreement, when it was only intended as a guide to the Java API. There are a lot of details in safe construction of such a protocol. More secure alternatives than to trying to roll this yourself include the various complete protocols listed at the end of the article. With that said, we’ll get back to the original article:
Diffie-Hellman key agreement (DH) is a way for two parties to agree on a symmetric secret key without explicitly communicating that secret key. As such, it provides a way for the parties to negotiate a shared AES cipher key or HMAC shared secret over a potentially insecure channel. It does not by itself provide authentication, however, so it is vulnerable to man-in-the-middle attacks without additional measures. There are several ways to provide these additional measures (e.g. signing the ephemeral public keys using a CA-issued certificate, or using a protocol like OTR), but we will not discuss them here, or go into the details of how the key agreement works. Java provides support out-of-the-box for both original discrete log DH and elliptic curve (ECDH) key agreement protocols, although the latter may not be supported on all JREs. ECDH should be preferred for any new applications as it provides significantly improved security for reasonable key sizes.
As is often the case in Java, the use of these classes can be a bit convoluted. Here we demonstrate simple Java code for ECDH key agreement on the command line. We only demonstrate ephemeral key agreement, in which the two parties generate unique public/private key pairs at the start of the protocol and throw them away once the shared secret has been negotiated. This can form the basis for perfect forward secrecy.
WARNING: the code here is not a complete security protocol and should be used for reference on the Java API only.