You should probably disregard this advice. Instead:
* Do not use passwords as API authentication. The user of an API is a computer program, not a human. Issue single-purpose random credentials for API access.]
* Make sure that your API is accessible only over HTTPS; test the API endpoint to ensure requests aren't honored over unencrypted HTTP.
* Use the simplest API authentication mechanism that (a) works and (b) you understand. You should probably degrade the capabilities of your API before you try to adopt an authentication system that you don't completely, fully, absolutely grok.
This means, if you don't need delegation right away (for instance, if you don't have 3rd party applications making calls to your API on behalf of your users), don't bother with anything like OAuth.
Contrary to the implications in this article, HTTP Basic Auth is not less secure than OAuth if your endpoints all require HTTPS. If you allow HTTP calls to an API, you have bigger problems than your auth token scheme.
Semantic mismatches between API endpoints and web UI endpoints (ie, semantic differences between API auth and cookie-based authentication) are a classic, pervasive, hard- to- eradicate source of serious security flaws. Don't make things any more complicated than they need to be.
[Late edit: I didn't think I needed to make the first point, but I clearly did need to. Sorry.]
> Semantic mismatches between API endpoints and web UI endpoints (ie, semantic differences between API auth and cookie-based authentication) are a classic, pervasive, hard- to- eradicate source of serious security flaws. Don't make things any more complicated than they need to be.
Could you explain that one in more detail? I don't think I understand it.
The one true solution is client authenticated TLS. For the life of me I can't understand why 25 years later certificate management is still such a mess.
We don't need some perfect universally recognized root trust system to get started. Why doesn't the sign up process for authenticated API access routinely include the issuance of a certificate signed by the API owner?
Unlike for interactive users there's no expectation that a customer will be accessing an API from some random computer where he might not have access to his certificate store.
I agree it's awful to use, but what I don't understand is why no one has bothered to improve the tools.
It's one of those areas where the underlying tool (either a library like openssl or NSS, or an OS feature like SSPI) could do the hard work in one place and make it simple for downstream libraries to wrap the functionality.
In other, somewhat analogous, domains that happened, but for whatever reason not in this case.
Well, in this case (having gone back and actually checked the docs rather than relying on recollection roughly a decade old) you could store MD5(username:realm:password) rather than password, though that doesn't buy you all that much in a case like this where making the password a long, random string that's not reused is not only good practice but also easy and so likely to be general practice.
I expected something very different from this article based on the headline. Specifically, I completely disagree with the "leave it to industry standards" approach, as that doesn't help people understand what they should do, and more importantly, why.
Request/response protocols (well, many things) really break down into 5 top-level categories (some sources will say the 6th is Audit):
It's a far more interesting exercise to walk through what you would get from each solution. Basic-Auth over TLS, actually gets you quite a ways towards that goal (specifically, C-I-A (authentication)). Where that, and notably HMAC, fall over is non-repudiation because they're based on a shared-secret model; admittedly HMAC keys are better than passwords because you're not sending the secret on every request, but asymmetric crypto is preferred. Authorization that the
server system does is really out of scope in all of these protocols, so Basic-Auth over TLS doesn't really impact that one. It can be as simple as "caller = owner," or as full-featured as a security policy language  (full disclosure: I am the original author of ).
OAuth really doesn't differ that much besides specifically solving the delegated access and website SSO problem(s); but IMO it does so with an overly baroque protocol that has too many parts. The "long pole" of setting up such a system (that is, allowing 3rd party sites to act on behalf of my site's users) isn't the specifics of what my REST api
looks like, but really it's all the "governance" of user decisions, and more-over, key management (in all these cases, key management is generally the hardest or almost hardest problem).
While it can be debated whether it was right or wrong, we (Joyent) released an open-source spec to solve straight up authentication of REST requests using SSH keys . At the end of the day, the user signs the Date header of requests with their private key (which by definition the server has never seen), and all requests must be over TLS. Disregarding Authorization, this scheme gives you Confidentiality (TLS), Integrity (TLS), Authentication (Signature), Non-Repudiation (asymmetric signature), and adds a "poor man's nonce,"
assuming you disallow requests where the clock skew of the date header is too large. And lastly, SSH solves a lot of key management problems for humans. Note: I didn't drop that reference to advocate for our specification here, but rather the security process you should think about when evaluating whether a protocol is secure.
Mutual-Auth SSL/TLS is a royal PITA, and is basically guaranteed to cause you grief. The client compatibility matrix might as well be considered an NP hard problem to assure yourself coverage, and failure modes of the different browsers/SDKs all differ. As a REST API should
have maximum accessibility to clients (i.e., don't wed yourself to any one language/sdk), this is pretty much a non-starter.
> Mutual-Auth SSL/TLS is a royal PITA, and is basically guaranteed to cause you grief. The client compatibility matrix might as well be considered an NP hard problem to assure yourself coverage, and failure modes of the different browsers/SDKs all differ. As a REST API should have maximum accessibility to clients (i.e., don't wed yourself to any one language/sdk), this is pretty much a non-starter.
Given the inconsistent support for mutual-auth TLS, which along with its direct ancestors has existed for a decade and a half, what makes you think they'll be widespread and correct support for SAS?
If by SAS you mean SSH, then really it's that there's a much smaller client surface area (really, openssh), and the fallback is always "use a PKCS#8 private key", which is much more uniform than SSL. SSL mixes in x509, and protocol, and UX.
That said, you don't have to agree with it. My macro point was really assessing these things against basic security threat modeling, not whether or not you agree with our choices of using SSH.
Well it's more than SSH. The client library needs to to implement the headers and do the date signing. Sure it's easier than TLS, but the libraries suffers more from a lack of dedicated interest than insane complexity I would think.
But I do agree with your basic point that too many people stop at authentication, instead of considering the full range of concerns.
UUID v4 is, in fact, random and "alphanumeric" in the sense that it's hex.
That being said, I have a few other issues with their wording as well. They should just say "we have a custom HMAC-based authentication scheme for our REST API". Also, it took me about 3 days to realize HMAC over SSL/TLS is about as secure and easy as you can get for most any language -- If you can send HTTP requests, you can probably do HMAC. You can add further safety by making expiring private keys for HMAC and other things, although my use cases are based on long running (weeklong+) batch computations, and not end users. (i.e. initial distribution of an expiring private key for HMAC over SSL, reauthentication schemes, etc...
> Best practices say to encrypt your passwords in the database to limit a potential data breach. This increases overhead for each request when authenticating a user. Unique API keys authentication skips the hashing step and therefore speeds up your calls.
Wait, why can you can skip the hashing step and still be secure? Because hashing is only neccesary if you call it a 'password', but not if you use it in the same way but call it an 'api key' instead?
I guess it depends on the purpose for hashing. If it's just about 'data breaches', then maybe it doesn't matter if your api keys get out... because they at least won't grant access to any _other_ systems, since they weren't manually entered by users and re-used accross systems. Is that what you're thinking?
But don't you still want to avoid a data breach like this for your _own_ service?
And, I think, isn't the other reason hashed passwords are used, to make it harder to do a brute force attack? Ie, it's quite intentional and deliberate that it increases request overhead. And doesn't this still apply to an api key, possibly even MORE so when you have a single api key instead of a username/pass combo?
If you lose the database for your own service, it does not matter how you authenticate to that service. Attackers have already fatally bypassed your credential system. At the same time, API credentials should by definition be single-use.
HMAC authentication requires both the client and the server to have a shared secret (or more
likely a derived key based on a shared secret). The secret cannot be saved as a one-way hash
(as might be common for a password). So you couldn't use BCrypt or SCrypt to hash the shared
(or derived) secret since the server would never be able to acquire the value to calculate the
You can still encrypt the secret, e.g. using AES 256 bit encryption with secure random Initialization
Vectors and rolling keys. This too is not easily 'brute forceable', but is very fast to decrypt compared to a BCrypt comparison (key storage should also be in a separate location than the main data store).
Obscurity can be a fine part of being secure. You have to think carefully about what happens when it fails, and realize that automated bots will tend to find things humans would consider obscure, but don't let the usually-good heuristic become a straightjacket.
Stormpath's custom scheme is very similar to Amazon's.
But per the blog article, you'd only want to do this if you are willing to support client libraries/sdks that implement it as well. No one wants to spend the time to implement non-standard custom HMAC algorithms.
Thanks for the pointer, I was looking for a sane explanation of the Amazon security and signature algorithm.
I have to disagree with the other comments here regarding the client library. I think that given the precedence of Amazon API, given that people understand how to sign APIs like amazon, this method will be accepted even without a client library.
Being a Java / Scala dev, I prefer that an API provider allow me to select the HTTP client libraries to use and prevent from forcing me to use a specific library & version via SDK transitive dependencies.
The advantage to OAuth over basic-auth is that you can delegate the credentials, which is to say you can set up your system so that users can give a limited-use credential to a 3rd party application to perform API calls on their behalf.
Some very large services have abused OAuth to "delegate" credentials to mobile devices, which has set up the expectation among developers that OAuth is the "sophisticated" way of doing all-around credential management of any sort. Not so. If you don't have delegation to third parties, don't use OAuth.
Most applications do not need delegation.
If you need delegation, there are simpler ways to do it than OAuth that won't meaningfully sacrifice the security of your controls. At the same time, using an OAuth solution you don't fully understand (for instance, using OAuth through a high-level library that hides the details from you) can damage the integrity of your whole application by creating new classes of mistakes for you to make.
There are many types of digest authentication - OAuth1.0a and Amazon's and Stormpath's custom schemes are examples. Browser-specific digest authentication wasn't covered however since the article was about REST APIs and most REST clients are not browsers.
I gathered as much. But in practice, how often do you see RFC 2617 Digest authc used in non-browser scenarios? (I'm genuinely curious. I haven't seen it used much at all outside of web browsers, so I'm curious what others may have come across).
I've written Atom Publishing Protocol servers that use it.
It's not badly-suited for non-browser tasks (although yes, SSL and Basic is much simpler - if you don't mind paying for the certificate). It's unusual, but it's pretty unusual to use it (or Basic) for web browsers these days, too.