Cookies are bad for you: Improving web application security
Most web applications today use browser cookies to keep a user logged in while she is using the application. Cookies are a decades-old device and they do not stand up well to security threats that have emerged on the modern web. In particular, cookies are vulnerable to cross-site request forgery. Web applications can by made more secure by using OAuth for session authentication.
When a user logs into a web application the application server sets a cookie value that is picked up by the user's browser. The browser includes the same cookie value in every request sent to the same host until the cookie expires. When the application server receives a request it can check whether the cookies attached to it contain a value that identifies a specific user. If such a cookie value exists then the server can consider the request to be authenticated.
There are many types of attacks that can be performed against a web application. Three that specifically target authentication between the browser and the server are man-in-the-middle (MITM), cross-site request forgery (CSRF), and cross-site scripting (XSS). Plain cookie authentication is vulnerable to all three.
In a MITM attack the attacker is in a position to watch traffic that passes between some user's browser and an application server. If that traffic is not encrypted the attacker could steal private information. One of the most dangerous things that an attacker can do in this position is to hijack the user's session by reading cookie data from an HTTP request and including that cookie data in the attacker's own requests to the same server. This is a form of privilege escalation attack. Using this technique an attacker can convince an application server that the attacker is actually the user who originally submitted a given cookie. Thus the attacker gains access to all of the user's protected resources.
Last year a Firefox extension called Firesheep made some waves when it was released. The purpose of Firesheep was to raise awareness of the danger of MITM attacks. Most web applications, at that time and today, use cookie authentication without an encrypted connection between browser and server. Firesheep makes it easy to spy on anybody who is using well known applications like Facebook and Twitter on a public network. With the click of a button you can perform a MITM attack yourself, steal someone's cookies, and gain access to that person's Facebook account.
MITM attacks can be effectively blocked by using HTTPS to encrypt any traffic that contains sensitive information or authentication credentials. When using HTTPS you will almost certainly want to set the "secure" flag on any cookies used for authentication. That flag prevents the browser from transmitting cookies over an unencrypted connection.
More and more web applications are offering HTTPS - often as on opt-in setting. Any web site that requires a login should offer HTTPS - and ideally it should be enabled by default.
CSRF attacks authentication indirectly. A malicious web page can trick a browser into making cross-domain requests to another web site. If a user visiting the malicious page is already logged in to that web site then the malicious page can access the site resources as though it were logged in as the unsuspecting user. For example, if a malicious page can trick the browser into making POST requests to a microblogging site it can post updates with spam links that appear to have been written by the victim.
If you use Facebook you might have encountered attacks like this yourself. You see a post on a friend's wall with a button that says "Don't click the button!" When you click on it you are taken to another site and the same message ends up posted on your wall.
This works because the browser automatically sends cookies set on a given domain with every request made to that domain, regardless of where those requests originated. The browser has no way of knowing that the requests initiated by the malicious page are made without the user's knowledge.
The malicious page could create a cross-domain request by including an
image with a
src attribute pointing to a URL on the site that it is
trying to hack into. The URL does not have to be an image - the browser
will make a GET request to that URL and will discard the response when
it determines that the response is not image data. If that GET request
produced any side-effects, like posting a microblogging update, then the
malicious page has successfully performed an attack.
To make a cross-domain POST request the malicious site might include
a hidden HTML form with an
action attribute pointing at the site to be
without any interaction from the user. This is another case where the
attacker cannot read the response that comes back but can trigger some
action in the user's account.
The biggest problem with CSRF is that cookies provide absolutely no defense against this type of attack. If you are using cookie authentication you must also employ additional measures to protect against CSRF. The most basic precaution that you can take is to make sure that your application never performs any side-effects in response to GET requests.
To protect against cross-domain POST requests a commonly used option is to use an anti-forgery token that must be submitted with every POST, PUT, or DELETE request. The token is generally injected into the HTML code for forms in such a way that malicious code on another site does not have any way to access it.
Such multi-layered approaches to CSRF defense work but are a pain to implement. I know from experience that the stateful nature of anti-forgery tokens make them a constant source of bugs in Ajax-driven applications where users might submit several requests to the server without ever loading a new page. It is too easy for the client and server to get out of sync and to disagree about which anti-forgery tokens are fresh. And great care must be taken to include the anti-forgery feature in every form and Ajax call in an application or a security hole appears.
JSON obfuscation is easier to apply to every JSON response as a blanket policy thanks to server-side filters and client-side hooks, such as those in jQuery's Ajax stack. But then you are not really serving JSON - you are serving a JSON-like type with a proprietary wrapper. I find that I spend a lot of time instructing people on the existence of obfuscation, explaining why it is there, and explaining how to set up hooks to remove it on the client side.
By combining the "secure" and "httpOnly" flags and using HTTPS you can make your application authentication proof against MITM attacks and against some XSS attacks. But there is nothing that will make cookie authentication resistant to CSRF attacks. The only way to protect against CSRF is to apply additional security measures. Often multiple measures are required to combat different possible CSRF vectors. And those measures are not always simple or transparent.
The way to make the web a safer place is to switch to authentication mechanisms that provide strong protection against CSRF at the most basic level. The key is to choose a mechanism that is controlled by the web application, not the browser. The web browser has no way of distinguishing legitimate requests from forged ones - it will attach cookies to both. On the other hand, application code can be written to be smarter.
There are many authentication schemes that would work well. I lean toward OAuth 2.0. OAuth has some nice advantages: it is standardized; there are numerous server implementations; and the simplest form of the OAuth 2.0 draft specification is pretty easy to implement.
In a traditional OAuth setup there are three parties: the authorization server / resource server, the client and the resource owner. Through a series of steps the resource owner, typically a user working through a web browser, submits a password to the authorization server and the authorization server issues an access token to the client. You can read more about the OAuth protocol flow on the OAuth 2.0 web site.
When applying OAuth to session authentication the picture becomes simpler: the browser acts as both the resource owner and the client; so some of the indirection of three-legged OAuth can be skipped. Instead, a web application can use a protocol flow that the OAuth 2.0 specification calls Resource Owner Password Credentials in which the user enters her password into a login form, the password is submitted to the application server directly, and the server responds to that request with an access token. You can think of this as "two-legged" OAuth.
In both the two- and three-legged flows requests are signed by adding an "Authorization" header with one of two possible formats. In the bearer scheme the authorization header value is just the access token that was given to the client. For example:
The HMAC scheme is a bit more complicated: in that case the client is given an access token id in addition to the token itself and the authorization header includes the token id and an HMAC-signed hash of the request URL, the request method, a nonce, and possibly a nested hash of the request body. The OAuth access token is used as the key in the HMAC algorithm.
The advantage of the HMAC scheme is that it can provide some protection against MITM attacks even if signed requests are not encrypted with HTTPS.
I propose a design in which the browser submits credentials from a login form to the server via XHR, gets an access token back, and uses that access token to sign subsequent requests. Full page requests and form posts are difficult to sign with OAuth - hyperlinks and form tags do not provide a way to specify an "Authorization" header. So OAuth-signed requests would probably be limited to XHR. The browser could store the OAuth access token in a persistent client-side store to give the user an experience that is indistinguishable from a cookie-based application - but that is more secure.
There are a couple of options for dealing with full page loads. One possibility is to not require any authentication for requests for HTML pages and to design your application so that HTML responses do not include any protected information. Such an application would serve pages as skeletons, with empty areas that to be filled in with dynamic and protected content after page load using Ajax. The dynamic responses could be HTML fragments that are protected by OAuth, or they could be JSON responses that are rendered as HTML using client-side templates.
Facebook uses a process like this which they call BigPipe. Facebook's rationale for BigPipe is actually performance, not security. In my opinion the BigPipe approach gives a best-of-both-worlds blend of performance and security. Plus, it lets you put caching headers on full page responses, even in apps with lots of dynamic content.
A downside of BigPipe is that content that is loaded via Ajax generally cannot be indexed by search engines. Google's recently published specification for making Ajax applications crawlable may provide a solution to that problem. Or you might choose to use the BigPipe approach everywhere in your application except for publicly accessible pieces of content.
Another way to handle full page loads would be to continue using cookie authentication for HTML resources. HTML responses are less vulnerable to CSRF snooping than JSON because HTML is not executable in script tags. In this case you should still require OAuth signing on requests for JSON resources and on any requests that could produce side-effects. But allowing cookie authentication on non-side-effect-producing GET requests for HTML resources should be safe.
A drawback is that you lose the XSS protection that the "httpOnly" cookie flag provides with cookie authentication. An application that uses OAuth will have to use other methods to block XSS. But in my opinion there are better options for dealing with XSS than there are for dealing with CSRF. By consistently sanitizing user-generated content you can effectively block XSS at the presentation layer of your application. That would be necessary anyway, since "httpOnly" only prevents XSS-based privilege escalation attacks and by itself does not prevent other XSS shenanigans.
To track a session using OAuth applications will need some way to store access tokens for the duration of a user's session. There are various ways to do that:
localStorage can be used to store a token so that is persistent even if the user closes and re-opens the browser. Data stored in localStorage is available to all windows on the same domain. You will probably want to include a hook to clear local storage when the user logs out of your application.
sessionStorage works like localStorage, except that data is only accessible from the same window that stored it and the whole store for a given window is wiped when the user closes that window. So the user does not have to log in again if she goes to another page; but she does have to log in again if she opens your app in a new window.
sessionStorage can be a more secure option than localStorage - especially on a shared or a public computer. If you decide to use a storage option that does not expire automatically when the browser is closed I suggest including a "remember me" checkbox in your login form and using sessionStorage instead when the user does not check that box.
Although I have been arguing that cookies are not the best option for authentication, storing an access token in a cookie works just fine. The key is that the server should not consider the cookie to be sufficient for authentication. Instead it should require that the access token be copied from the cookie value into an OAuth header.
A nice advantage of the cookie option is that users have been trained that they can delete cookies to reset a session. On the other hand, most users do not know about localStorage and most browsers do not provide an obvious way to clear localStorage. So the cookie option is likely to conform best to user expectations. Cookies can also be configured to expire when the browser is closed or to persist for a long period of time.
There is a good summary of client-side storage implementations and how to use them on Dive Into HTML5. Or if you want a pre-built solution that avoids most cross-browser headaches you can use PersistJS or a similar tool.
You should consider your target audience, your application requirements, and your security needs and decide for yourself whether dropping the noscript option is the right choice for your application.
No security protocol is bulletproof. Do lots of research and use common sense whenever you are working on an application that needs to be secure.
Other images used in diagrams are from the Open Clip Art Library and are in the public domain.