Comments:"cortesi - How mitmproxy works"
URL:http://corte.si/posts/code/mitmproxy/howitworks/index.html
I started work on mitmproxy because I was frustrated with the available interception tools. I had a long list of minor complaints - they were insufficiently flexible, not programmable enough, mostly written in Java (a language I don't enjoy), and so forth. My most serious problem, though, was opacity. The best tools were all closed source and commercial. SSL interception is a complicated and delicate process, and after a certain point, not understanding precisely what your proxy is doing just doesn't fly.
The text below is now part of the official documentation of mitmproxy. It's a detailed description of mitmproxy's interception process, and is more or less the overview document I wish I had when I first started the project. I proceed by example, starting with the simplest unencrypted explicit proxying, and working up to the most complicated interaction - transparent proxying of SSL-protected traffic in the presence ofSNI.
Configuring the client to use mitmproxy as an explicit proxy is the simplest and most reliable way to intercept traffic. The proxy protocol is codified in the HTTP RFC, so the behaviour of both the client and the server is well defined, and usually reliable. In the simplest possible interaction with mitmproxy, a client connects directly to the proxy and makes a request that looks like this:
GET http://example.com/index.html HTTP/1.1
This is a proxy GET request - an extended form of the vanilla HTTP GET request that includes a schema and host specification, and it includes all the information mitmproxy needs to relay the request upstream.
The process for an explicitly proxied HTTPS connection is quite different. The client connects to the proxy and makes a request that looks like this:
CONNECT example.com:443 HTTP/1.1
A conventional proxy can neither view nor manipulate an SSL-encrypted data stream, so a CONNECT request simply asks the proxy to open a pipe between the client and server. The proxy here is just a facilitator - it blindly forwards data in both directions without knowing anything about the contents. The negotiation of the SSL connection happens over this pipe, and the subsequent flow of requests and responses are completely opaque to the proxy.
The MITM in mitmproxy
This is where mitmproxy's fundamental trick comes into play. The MITM in its name stands for Man-In-The-Middle - a reference to the process we use to intercept and interfere with these theoretically opaque data streams. The basic idea is to pretend to be the server to the client, and pretend to be the client to the server, while we sit in the middle decoding traffic from both sides. The tricky part is that the Certificate Authority system is designed to prevent exactly this attack, by allowing a trusted third-party to cryptographically sign a server's SSL certificates to verify that they are legit. If this signature doesn't match or is from a non-trusted party, a secure client will simply drop the connection and refuse to proceed. Despite the many shortcomings of the CA system as it exists today, this is usually fatal to attempts to MITM an SSL connection for analysis. Our answer to this conundrum is to become a trusted Certificate Authority ourselves. Mitmproxy includes a full CA implementation that generates interception certificates on the fly. To get the client to trust these certificates, we register mitmproxy as a trusted CA with the device manually.
Complication 1: What's the remote hostname?
To proceed with this plan, we need to know the domain name to use in the interception certificate - the client will verify that the certificate is for the domain it's connecting to, and abort if this is not the case. At first blush, it seems that the CONNECT request above gives us all we need - in this example, both of these values are "example.com". But what if the client had initiated the connection as follows:
CONNECT 10.1.1.1:443 HTTP/1.1
Using the IP address is perfectly legitimate because it gives us enough information to initiate the pipe, even though it doesn't reveal the remote hostname.
Mitmproxy has a cunning mechanism that smooths this over - upstream certificate sniffing. As soon as we see the CONNECT request, we pause the client part of the conversation, and initiate a simultaneous connection to the server. We complete the SSL handshake with the server, and inspect the certificates it used. Now, we use the Common Name in the upstream SSL certificates to generate the dummy certificate for the client. Voila, we have the correct hostname to present to the client, even if it was never specified.
Complication 2: Subject Alternative Name
Enter the next complication. Sometimes, the certificate Common Name is not, in fact, the hostname that the client is connecting to. This is because of the optional Subject Alternative Name field in the SSL certificate that allows an arbitrary number of alternative domains to be specified. If the expected domain matches any of these, the client will proceed, even though the domain doesn't match the certificate Common Name. The answer here is simple: when extract the CN from the upstream cert, we also extract the SANs, and add them to the generated dummy certificate.
Complication 3: Server Name Indication
One of the big limitations of vanilla SSL is that each certificate requires its own IP address. This means that you couldn't do virtual hosting where multiple domains with independent certificates share the same IP address. In a world with a rapidly shrinking IPv4 address pool this is a problem, and we have a solution in the form of the Server Name Indication extension to the SSL and TLS protocols. This lets the client specify the remote server name at the start of the SSL handshake, which then lets the server select the right certificate to complete the process.
SNI breaks our upstream certificate sniffing process, because when we connect without using SNI, we get served a default certificate that may have nothing to do with the certificate expected by the client. The solution is another tricky complication to the client connection process. After the client connects, we allow the SSL handshake to continue until just after the SNI value has been passed to us. Now we can pause the conversation, and initiate an upstream connection using the correct SNI value, which then serves us the correct upstream certificate, from which we can extract the expected CN and SANs.
There's another wrinkle here. Due to a limitation of the SSL library mitmproxy uses, we can't detect that a connection hasn't sent an SNI request until it's too late for upstream certificate sniffing. In practice, we therefore make a vanilla SSL connection upstream to sniff non-SNI certificates, and then discard the connection if the client sends an SNI notification. If you're watching your traffic with a packet sniffer, you'll see two connections to the server when an SNI request is made, the first of which is immediately closed after the SSL handshake. Luckily, this is almost never an issue in practice.
Putting it all together
Lets put all of this together into the complete explicitly proxied HTTPS flow.
When a transparent proxy is used, the HTTP/S connection is redirected into a proxy at the network layer, without any client configuration being required. This makes transparent proxying ideal for those situations where you can't change client behaviour - proxy-oblivious Android applications being a common example.
To achieve this, we need to introduce two extra components. The first is a redirection mechanism that transparently reroutes a TCP connection destined for a server on the Internet to a listening proxy server. This usually takes the form of a firewall on the same host as the proxy server -iptables on Linux orpf on OSX. Once the client has initiated the connection, it makes a vanilla HTTP request, which might look something like this:
GET /index.html HTTP/1.1
Note that this request differs from the explicit proxy variation, in that it omits the scheme and hostname. How, then, do we know which upstream host to forward the request to? The routing mechanism that has performed the redirection keeps track of the original destination for us. Each routing mechanism has a different way of exposing this data, so this introduces the second component required for working transparent proxying: a host module that knows how to retrieve the original destination address from the router. In mitmproxy, this takes the form of a built-in set ofmodules that know how to talk to each platform's redirection mechanism. Once we have this information, the process is fairly straight-forward.
The first step is to determine whether we should treat an incoming connection as HTTPS. The mechanism for doing this is simple - we use the routing mechanism to find out what the original destination port is. By default, we treat all traffic destined for ports 443 and 8443 as SSL.
From here, the process is a merger of the methods we've described for transparently proxying HTTP, and explicitly proxying HTTPS. We use the routing mechanism to establish the upstream server address, and then proceed as for explicit HTTPS connections to establish the CN and SANs, and cope with SNI.