Best practice for syncing private keys between server instances - rest

I'm working on a project that uses jwt (json web tokens) to check whether a user is authenticated with the system.
The services are completely stateless, so I can easily scale from one instance to a few dozens. There is only one issue. To be able to validate the jwt on every instance, I need to synchronize the symmetric secret key between the instances.
Of course I could build a system that uses some sort of service discovery and exchanges the symmetric key by encrypting it with an asymmetrical encryption. This just doesn't feel like a proper method.
I could set a fixed key on every instance but changing the key, which effectively invalidates all the previously created tokens, would be a pain.
I don't want to set a master instante every other instance is relying on.
My third idea was using an asymmetrical approach and letting each instance create their own key pair. By writing the issuer in the token, the instance that gets called could retrieve the public key from the instance that issued the token and validate the token this way.
It's not like I don't have any ideas. I just don't really like my ideas so far.
Do you have any best practices to solve such a problem?

Related

JWT key rotation

I have been thinking and searching many different approaches but I am not sure what is the best solution to solve this problem.
Imagine you have 50 microservices that communicate via symmetric JWT - shared key. At the moment the jwt is stored on each microservice (server or virtual host).
What i would like to do is to use aws secret manager to manager the JWT key and to eventually rotate the key.
Obviously i can rotate the key easily but i want to cache the JWT token (with the aws library for python) so that i do not make too many calls to the services (yes, each call to each web service requirs a JWT).
Obviously, if i cache the key for a certain amount of time, there will be cases where the key will be out of sync on some servers, so how would you move forward to be able to rotate secret in aws secret service (keeping in mind that i use caching to avoid slowdown in server interaction and cost)?
Any suggestions?
JWKS use asymmetric priv/pub keys kept in secret manager, expose pub key on the auth server (via url) for other services to use. Auth server either periodically checks for new keys or you ping it to refresh after rotating the keys.

JWT architecture for multiple users?

Most examples always take into consideration only one users using the system in the JWT/Flask tutorials. I want to understand this at a multi-user level but cannot find the correct resources.
Lets say we have the following secret key:
app.config['SECRET_KEY'] = 'randomkey'
Two Questions:
Would this key be the same for every user? if so, wouldn't this provide a security risk because if the key was stolen, anyone would have access to do whatever they want?
If it is not the same, how is the key stored on the server-side so it could be authenticated when requesting information? Would it be stored within the user's table under current token or something?
In this case, that key is the JWT signing key It could also be different from flask's secret key setting (see flask docs). It's not used for encryption so it's not intended to be a shared secret between server and users. Its role is to provide the server with a proof that the JWT content was generated by the server itself: it's a proof of integrity.
Having knowledge of that key means having the right to issue JWTs on behalf of the application, attackers could impersonate servers or make requests with some modified claims, eg pretending to be other users. This means that these keys are quite sensible from security perspective
Turns out that 1 app : 1 key, with some remarks
This key should theorically never change: if at time T1 KEY=x, a user could login in and receive a JWT signed with KEY=x. at T2 KEY=y, the user will invoke some API using the previous JWT and the server will try to verify(signature(payload , x) , y). So every user will be logged out automatically
Despite 1. It would be good to rotate the key. In this case the authentication system should save a buffer of old keys and use them to validate oldest JWT. Since a JWT should be short-lived, it could be useful set a rotation time grater than JWT expire time and just keep the last used key along with the current
This key is a secret and should be managed exactly as other secrets. Aside from terrible approaches like leaving it plaintext in the code/config, there are secret managers from cloud providers, or kubernetes secrets if you use the latter, as well as secret managers from configuration management tools (salt, ansible) or Hashicorp's Vault that is a specialized storage engine for sensible data. Anyway, it's more infra/security team concern if you are in a structured organization

What benefit does JWT provide?

I have implemented JWT based security in a test Core Web API REST project, it is working fine but I am not sure that I see the benefit of this. The web says JWT is good because it's lightweight and can be used to verify that the source of data but in my implementation:
The client first provides a username and password to authenticate
If user + pwd is ok the a token is returned and every subsequent call to the api uses that jwt token (instead of the username and password) to authenticate.
This is fine but why not just use the username + password on every call to the api (and skip the complication of managing the token)?
In fact in my case there's additional complications because I now have to factor in an expiry date (of the token) that resides outside of my system.
Can someone explain what I'm missing here?
One of the main benefits and motivations for using JWT is that it allows your server side application to push all session state information outside of the application. That is, in a theoretical limit, a JWT implementation is actually stateless.
To directly answer your question, we can compare the workflows for what happens when username/password is submitted in every request versus submitting a JWT.
First, a JWT contains a claims section, which is typically written by the issuer of the token, i.e. the server side application. One of the fields is called exp, and contains the expiry time of the token. One property of JWT is that it is not possible for the user to tamper with them. This is enforced via a checksum, which would change if any part of the JWT changes. Taken together, this means that the user cannot alter the expiry time (or any other claim), and the server can implicitly trust this time. When the user submits a request with a JWT, in theory all the server has to do is just check exp to see if the token still be valid. That is, the session state actually lives outside the application, at least in theory.
In contrast, when the user submits a username/password each time, the server has no way of knowing what to do just based on that information. Rather, the server has to maintain the session state itself, and this can be costly both in terms of memory and performance.
In practice, JWT is never completely stateless, but, using a good implementation, it is usually possible to get the memory footprint very small, requiring only a bit of space in a cache (e.g. Redis or a similar tool).

JWT and one-time tokens?

I'm in the process of rolling my own JWT token auth, however, I would really like it to be a one time-token – so once it's used, the server generates a new token and the client will have to use that token during the next request/call.
However, it has come to my understanding that JWT is supposed to be 'stateless' – but with the approach of a one time token, I guess I would need to somehow store the valid tokens, since the token will be refreshed once it's used. Or is there any way to avoid storing a value on the server, and still be able to create one-time tokens?
The two main reasons for why I don't want to store any value is first of all scalability (sure, I could have cache-server inbetween to store the values, but it would be nice if that wasn't required), secondly, JWT is supposed to be stateless from my understanding, which it wouldn't be if I need to store a value on the server to be able to validate the token.
Any ideas?
Use the user's current password's hash for signing the JWT token, in this way all tokens generated before a successful password change would get invalidated the next time. I got the idea from here https://www.jbspeakr.cc/howto-single-use-jwt/.
Solutions exist, of course.
As with any distributed system (you mentioned scalability) you have to choose between availability and consistence.
You choose availability. In this case you could maintain a list of already-used tokens that you replicate in a eventually consistent manner between all the endpoints. For example when a token is used the respective endpoint send that token to the other endpoints in the backgound. There is however a (short) time frame when that token can be used a second time by another endpoint until that endpoint is updated.
You choose consistency (you won't allow a token to be used multiple times whatsoever). In this case you use a central database with already-used tokens and you check that database everytime you need to perform an action. Scalability? You could use sharding on the token and have n databases, each one being responsible for a tokens subset.
It depends on your business what solution fits best.
Not really no, a JWT token is valid if it hasn't expired and the signature is correct, commonly people will keep a DB of blacklisted tokens which are usually ones where people have logged out etc.
The only sensible way I can think of is give them a short expiry time and maintain a list of tokens that have already been used, you'd then periodically remove the ones that subsequently expire from the DB.
There are actually some DB's that have a TTL on records (dynamoDB, mongodb) so you'd just put the tokens in and set a TTL for when the token expires.
Update 2022
Just to be clear JWT tokens AREN'T stateless they have claims that, as long as they're signed by the right private key - give you a stateful piece of data that can be reissued by your API to reflect the current state of the user.
You'd just need to handle token re-issue on the consumer.
Like others have mentioned, it depends on your business case. Password resets links can be like mentioned on https://www.jbspeakr.cc/howto-single-use-jwt/.
If you have the Single-Use & Single-Auth scenario, where you might want to invalidate any previously used and unused token, you can store a single nonce and update it on every new token request and also when its used.

How to securely communicate with server?

I'm building a solution consisting of an app and a server. Server provides some methods (json) and the app uses them. My aim is to make those API methods inaccessible to other clients. What is the best way to do so?
Should I take a look at certificates (to sign every outgoing request)? If yes, where do I start and what is the performance impact of doing so?
What are alternatives?
Put another way, you need a way to distinguish a valid client's request from an invalid client's request. That means the client needs to present credentials that demonstrate the request comes from a valid source.
SSL certificates are an excellent way to assert identity that can be validated. The validity of an SSL certificate can be confirmed if the certificate contains a valid signature created by another certificate known to be secure, a root cert. As noted in other answers an embedded certificate won't do the job because that certificate can be compromised by dissecting the app. Once it is compromised, you can't accept any requests presenting it, locking out all your users.
Instead of one embedded app cert, you need to issue a separate certificate to each valid user. To do that, you need to set up (or outsource to) a Certificate Authority and issue individual, signed certificates to valid clients. Some of these certificate will be compromised by the user -- either because they were hacked, careless or intentionally trying to defraud your service. You'll need to watch for these stolen certificates, place them on a certificate revocation list (CRL) and refuse service to these compromised certificates. Any web server is able to refuse a connection based on a CRL.
This doesn't solve the security issues, it just moves them out of the app. It is still possible for someone to create what appears to be a valid certificate through social engineering or by stealing your root certificate and manufacturing new signed certificates. (These are problems all PKI providers face.)
There will be a performance hit. How much of a hit depends on the number of requests from the app. The iPhone NSURLConnection class provides support for SSL client certificates and client certificates can be installed in the phone from an e-mail or authenticated web request. Managing the infrastructure to support the client certs will require more effort than coding it into the app.
Incidentally, voting down any answer you don't like creates a chilling effect in the community. You're not nearly as likely to get advice -- good or bad -- if you're going to take a whack at everyone's reputation score.
I will now freely admit that it's an interesting question, but I have no idea how it could be done.
Original answer:
Interesting question. Assuming people can't reverse-engineer the iPhone app, the only solution that comes to mind would be to sign requests with a public key, or some other secret known only to the application. By that, I mean adding an extra argument to every API call that is a hash of the destination URL and other arguments combined with a secret known only to your server and application.
To expand upon this: suppose your API call has arguments foo, bar and qux. I would add a signature argument, the value of which could be something as simple as sorting the other arguments by name, concatenating them with their values, adding a secret, and hashing the lot. Then on the server side, I would do the same thing (excepting the signature argument) and check that the hash matches the one we were given in the request.
Consider authenticated HTTP.
For a cheaper alternative, there's shared secret/hash scheme. The client and the server have a shared secret string of text. Upon request, the client hashes together (using MD5, or SHA1, or SHA something else - you choose) the request fields and the secret. The hash value is attached to the request - say, as another POST field.
The server does the same operation with the request and with its copy of the secret, then compares the hash values. If they don't match - service denied.
For added security, you may encrypt the hash with a RSA public key. The client has the public key, the server keeps the private key. The server decrypts the hash with the private key, then the same. I did that with a C++ WinMobile client and a PHP-based service - works like a charm. No experience with crypto on iPhone, though.
UPDATE: now that I think of it, if we assume that the attacker has complete control over the client (ahem jailbroken iPhone and a debugger), the problem, as formulated above, is not solvable in theory. After all, the attacker might use your bits to access the service. Reverse-engineer the executable, find the relevant functions and call them with desired data. Build some global state, if necessary. Alternatively, they can automate your UI, screen scraper style. Such is the sad state of affairs.