Introducing TAuth: Why OAuth 2.0 is bad for banking APIs and how we're fixing it

This week we released our authorisation flow making it possible for you to go from building apps that talk to your bank account, to building apps that can talk to any bank account. This is huge. Check out this SMS bot (how on trend) I hacked up yesterday morning. (and don't forget to join the beta wait list).

Getting to this point took longer than we expected. This is because there wasn't a good story for delegating authorisation for sensitive APIs. The most popular choice, OAuth 2.0 - which has been chosen by the Open Banking Working Group, BBVA, RBS, and Mondo - is also amongst the worst from a security perspective.

Teller provides an API for your bank account. The EU is forcing all European banks to expose account APIs with PSD II by end of 2017. These banks are disconcertingly converging around OAuth 2.0* without fully considering the impact on their customers, and something needs to be done before it's too late.

* One notable exception is the Open Bank Project. It is sticking with OAuth 1.0a precisely because OAuth 1.0a doesn't share the same security issues as OAuth 2.0.

Man-in-the-middle

One of the biggest problems with OAuth 2.0 is that it delegates all security concerns to TLS but only the client authenticates the server (via it's SSL certificate), the server does not authenticate the client. This means the server has no way of knowing who is actually sending the request. Is it a bona fide user, or is an attacker tampering with the request? When an attacker is able to insinuate themselves between a legitimate user and the server, it's called a man-in-the-middle (MITM) attack. It looks like this:

You're probably thinking "hang on, isn't this the point of SSL?" Yes it is, but there are a number of ways to present a bogus certificate and a client accept it. The most realistic threat is the client developer not properly verifying the server certificate, i.e. was it ultimately signed by a trusted certificate authority?

Pull requests to disable SSL certificate verification: more common than you would think.

— Tim Pope (@tpope) February 11, 2013

Unfortunately a large number of developers think that disabling SSL peer verification is the correct fix to a SSL path validation error. There are many more that will offer the same advice with the caveat that it introduces a security issue that < 100% of readers will consider. As an API provider with a duty of care to our users we can't simply hope developers on our platform don't do this.

Bearer tokens

Once a user authorises a client application to access its account, the application obtains a bearer token from the authorisation server. As the name suggests if you have possession of the bearer token then you are essentially the user. There is no cryptographic proof that the requesting client is the intended developer and not an attacker. If an attacker is able to successfully MITM a client it could have catastrophic implications for the user, e.g. an empty bank account, loans opened in their name, etc. OAuth 2.0 is simply a security car crash from a bank's perspective. They have no way to prove that an API transaction is bona fide, exposing them to unlimited liability.

For more information on OAuth 2.0 shortcomings see OAuth Bearer Tokens are a Terrible Idea and OAuth 2.0 and the Road to Hell by Eran Hammer the original primary author of OAuth 2.0 who formally removed his name from the standard, calling it "the biggest professional disappointment of [his] career."

Finding something better

Sitting down to design the solution to this problem I had two high-level goals:

From a security perspective I wanted:

(N.B. Non-repudiation is a de facto requirement of PSD II. If a bank can't prove an account owner authorised a transaction, they're liable for any losses incurred by the user.)

There are solutions to the bearer token problem like JWT tokens (RFC7523) but in most cases these rely on a shared secret which is used to computed a HMAC-based signature. Shared secrets mean no non-repudiation. Public key cryptography can be used with JWT tokens but they don't solve the problem of how the client will generate key pairs, demonstrate proof of possession of the private key, and enrol the public key with the API. Most importantly using JWT tokens make it basically impossible for you to experiment with an API using cURL. A major impediment to developer experience.

In an ideal world we'd have cryptographic proof of the client's identity without it having to leak through the application level (and stop us cURLing the API!). As I thought about it it became the clear the answer was hiding in plain sight: SSL client certificates.

Client SSL Authentication

A SSL handshake involving client certificates contains an extra message at the end of the handshake, the CertificateVerify message.

  Client                            Server

  ClientHello        -------->
                                    ServerHello
                                    Certificate*
                                    ServerKeyExchange*
                                    CertificateRequest*
                     <--------      ServerHelloDone
  Certificate*
  ClientKeyExchange
  CertificateVerify*
  [ChangeCipherSpec]
  Finished           -------->
                                 [ChangeCipherSpec]
                     <--------             Finished
  Application Data   <------->     Application Data

         Fig. 1 - Message flow for a full handshake

The client collects all the handshake messages and signs them with it's private key and sends the result to the server. The server then verifies the signature using the public key of the client certificate. If the signature can be verified with the public key, the server knows the client is in possession of the private key, and is therefore a bona fide user.

Let's look at this in the context of our original attack:

Introducing TAuth

TAuth is Client SSL Authentication + User Tokens + Great Tooling.

Client SSL authentication is often overlooked because of the poor UX of using client certificates in the browser and that generating certificates is a painful multistep process involving arcane OpenSSL CLI incantations. However as we're talking about API clients the browser UX point is irrelevant. As far as certificate generation goes, we can write better tools. These days it's possible to generate a key pair, a PKCS#10 certificate request, and sign it all in the browser. Thanks to WebCrypto the whole process is reduced to one click.

This is how Teller does it:

A private key and a SSL certificate signed by Teller generated in one click.

And this is what a request looks like with client certificates:

Let's recap what we've achieved here:

Token security

Notice in the above example. The Teller API accepts connections from clients without a client certificate. We do this because we provide developers with read-only personal access tokens for their own accounts if they want to quickly hack something up and not bother with provisioning certs. Now notice how the API does not accept the token presented, but accepts it when used with the client SSL certificate. TAuth bearer tokens are bound on the server side to a private key through an application. This means they are useless without the private key (which only the developer ever has) and therefore not sensitive. As matter of fact, here is one for my bank account:

Ruby client

You'll need the private key it's bound to for it to be of any use, and that has never left my laptop.

TAuth tokens do not expire (but can be revoked). OAuth 2.0 introduced the concept of time-limited tokens. Large internet companies found it useful for scaling purposes to issue self-encoded, encrypted tokens. The drawbacks are developers have to pay the complexity cost of refreshing tokens and most importantly tokens cannot be revoked, they're good until they expire. For bank account APIs this is an undesirable property, a token should be void as soon as the account owner wishes. TAuth checks the token revocation status at each request.

Given that we have no need for self-encoded tokens and that tokens are useless to anyone without the private key, we can consider them public and directly return them in the callback further simplifying things for the developer compared to OAuth without compromising the user's security.

Bonus: DDOS mitigation

TAuth can help mitigate layer 7 based DDOS attacks too. If the client does not present a valid client certificate the server can just choose to bounce the connection.

Conclusion

OAuth 2.0 is simply not fit for use with sensitive APIs and all the pieces needed to build something that is exist today. Aside from unbound bearer tokens, OAuth is too open-ended and complicated to get right for most developers. It's so bad it even has it's own threat model as a separate RFC to offer mitigations for the endless problems that can happen.

TAuth stands for Trusted Authentication, and it provides the best security for users while maintaining the highest possible quality developer experience. Less can go wrong when everything is simpler. If your bank has OAuth 2.0 in production you must ask yourself, do they really know what they're doing?

TAuth is available in production today for our existing beta users and we've already begun the work to make it an open standard we hope the industry adopts. If you're at a bank and want to offer your customers the security they deserve email me - sg at teller.io

Sign up for the beta waiting list at Teller.io.