Pwn2Own 2018: Safari + macOS

Safari RCE, sandbox escape, and LPE to kernel for macOS 10.13.3.

Usage

Install nasm and tornado:

brew install nasm
pip3 install tornado

Check config.py if you want to change the host or ports. Afterwards start the server with ./server.py and navigate to the shown URL.

Overview

This exploit chain uses three different bugs to go from JavaScript code running inside Safari to kernel-mode code execution:

1. An incorrect optimization in the DFG JIT compiler that can be used to cause a type confusion
2. Missing sandbox checks in launchd, allowing sandboxed processes to spawn arbitrary (non-sandboxed) processes
3. A logic bug in XNU, allowing a process to override the bootstrap port of its child processes, leading to an IPC MitM situation

The exploit chain is implemented in six stages, each located in its own subdirectory:

• stage0/: the WebKit exploit
• stage1/: the first stage payload written in assembly
• stage2/: the second stage payload to perform the sandbox escape
• stage3/: shell scripts to coordinate the remaining stages
• stage4/: a LPE to gain root
• stage5/: a LPE to gain kernel code execution
• libspc/: reimplementation of the XPC protocol, used by stages 2, 4, and 5

Every subdirectory (with the exception of libspc/) contains a file named make.py which, when executed, performs any kind of build command necessary and creates a list of files to be served by the webserver.

Stage 0

Goal: achieve shellcode execution inside the sandboxed WebContent process
Bug exploited: incorrect optimization in the DFG JIT compiler
See also this BlackHat talk

The DFG JIT compiler represents JavaScript code in its own intermediate representation (IR), the Data Flow Graph (DFG). Typically, one JavaScript expression will be translated to one or multiple IR instructions in this graph. In the case of a constructor function, the CreateThis instruction is emitted and is responsible for allocating the this object that is constructed by the function. As an example, the function function Consructor() {}, when called with new, would roughly be translated to

v0 = CreateThis
return v0

Looking at the AbstractInterpreter, we can see that the DFG JIT compiler assumes that the CreateThis operation will not result in any side effects besides a heap allocation. In fact, this code:

function Constructor(obj) {
    return obj.x;
}

will roughly be translated to the following DFG instructions: (Here, the StructureCheck was moved to the beginning of the function by the TypeCheckHoistingPhase).

StructureCheck(arg1);
v0 = CreateThis;
v1 = LoadOffset(arg1, OFFSET)
return v1;

However, that assumption is invalid, as the slow-path code for CreateThis can execute arbitrary JavaScript code in some cases. In particular, by using a Proxy around the actual function, the get trap for the "prototype" property will be called during the slow-path handler for CreateThis as it needs to fetch the prototype object for the constructed object:

function Constructor(obj) {
    return obj.x;
}

var handler = {
    get(target, propname) {
        /* run JS here, modify the structure of the argument object, etc. */
        return target[propname];
    },
};
var ConstructorProxy = new Proxy(Constructor, handler);

// Force JIT compilation of ConstructorProxy

As such, it is now possible to modify the Structure of an object without the JIT compiler performing a bailout.

This bug can be used to construct addrof and fakeobj primitives as follows:

addrof

We compile the code for the case of a JSArray with unboxed double elements, then, in the callback, transition to JSValue elements. Afterwards, the JIT code will load a JSValue from the array, but treat those bits as a double and return them to us. The following code will assign the address of leakme to the "address" property of the constructed object.

function InfoLeaker(a) {
    this.address = a[0];
}

var handler = {
    get(target, propname) {
        if (trigger)
            arg[0] = leakme;
        return target[propname];
    },
};
// ...

fakeobj

Here we essentially do it the other way around: we optimize code to store a double to an array with unboxed double elements, then again transition to JSValue elements in the callback. The code will continue to write our controlled double in unboxed form to the backing storage. When we later access that array element, it will treat those bits as a JSValue instead of a double. The following code will write the unboxed double address into the backing buffer of a which we can then read out as JSValue, allowing us to "inject" JSValues of our choosing into the engine.

function ObjFaker(a, address) {
    a[0] = address;
}

var handler = {
    get(target, propname) {
        if (trigger)
            arg[0] = {};
        return target[propname];
    },
};
// ...

As such we end up with the ability to write a double and treat is as JSObject pointer and vice versa. This can be exploited as described in attacking javascript engines.

The exploit first achieves arbitrary process memory read/write by faking a Float64Array, then searches for the JIT region (mapped RWX) and writes the stage1 shellcode there.

Stage 1

Goal: bootstrap stage 2 by writing a .dylib to disk and loading it via dlopen()

A short assembly payload which essentially does the following:

1. Call confstr(\_CS\_DARWIN\_USER\_TEMP\_DIR) to obtain a path to a writable directory
2. Create a new file named 'x.dylib' in the writable directory
3. Write the stage2 dylib into the newly created file
4. Load the dylib into the WebContent process through dlopen()

Stage 2

Goal: break out of the sandbox
Bug exploited: missing sandbox checks in launchd's "legacy_spawn" API
See also this talk

Launchd exposes the "legacy_spawn" RPC endpoint as routine 817 in subsystem 3. This API fails to validate whether the caller should be allowed to spawn processes and will just execve any binary on the system for the caller with controlled arguments. Since launchd is reachable through the bootstrap port, this makes it possible to escape from the sandbox.

The exploit essentially runs curl server/pwn.sh | bash and thus passes control to stage3.

Stage 3

Goal: pop calc and bootstrap the remaining stages

This executes open /Applications/Calculator.app and establishes a reverse shell, then fetches all files required for the remaining stages and runs the exploits.

Stage 4

Goal: gain root via a LPE exploit
Bug exploited: XNU bootstrap port MitM
See also this POC talk

In XNU, the task_set_special_port API allows callers to overwrite their bootstrap port, which is used to communicate with launchd. This port is inherited across forks: child processes will use the same bootstrap port as the parent. A security issue now arises if the child process is more privileged than the parent, as is the case for example with sudo (a setuid binary) or kextutil (having the "com.apple.rootless.kext-management" entitlement"). By overwriting the bootstrap port and forking a child processes, we can now gain a MitM position between our child and launchd (which our child expects to reach when sending messages to the bootstrap port). The child process will ask launchd to resolve various mach and XPC services. By resolving these services to other ports controlled by us, we can also gain a MitM position with arbitrary system services used by our child process. Exploitation then depends on how those services are used by the attacked program.

To gain root we target the sudo binary and intercept its communication with opendirectoryd, which is used by sudo to verify credentials. We modify the replies from opendirectoryd to make it look like our password was valid.

It appears that there was an attempt to fix this problem since libxpc (which performs the communication with launchd) verifies that the responses indeed come from a uid=0 and pid=1 (== launchd) process. However, these checks are insufficient. We can bypass them as follows to resolve opendirectoryd to our own port:

1. Register our own mach service (e.g. net.saelo.hax) with launchd using the bootstrap_register2 API
2. Intercept the service lookup request to launchd and replace the string com.apple.system.opendirectoryd.api with net.saelo.hax
3. Forward the request to launchd, but leave the original reply port in place, so launchd answers directly to the child process and the checks in libxpc in our child succeed

All that remains now (for a privilege escalation to root) is to forward the messages between opendirectoryd and sudo, but replace the authentication error reply with a success reply.

Stage 5

Goal: load a (self-signed) kernel extension
Bug exploited: XNU bootstrap port MitM

This exploits the same flaw as stage4, but this time targeting kextutil. We intercept the connection to com.apple.trustd and spoof the certificate chain, causing kextutil to think that our self-signed kext is actually signed directly by apple.

kextutil proceeds roughly as follows when asked to load a .kext from disk:

1. Verify the integrity of the .kext by checking all signatures against the provided certificate
2. Communicate with trustd to obtain the certificate chain and establish whether the root certificate is trusted
3. Verify that the root of the certificate chain is an apple certificate
4. Check whether the .kext is user-approved by talking to syspolicyd. However, if syspolicyd can not be reached, kextutil simply proceeds

This enables the following attack to load self-signed kernel extensions:

1. Create a .kext and sign it with a self-signed certificate
2. Run kextutil and resolve com.apple.trustd to our own service
3. Intercept messages to trustd and reply with a hardcoded certificate chain of an official apple .kext
4. Block communication with syspolicyd (e.g. by replacing com.apple.security.syspolicy.kext with net.saelo.lolno in service lookup requests to launchd)

kextutil will now load our kernel extension into the kernel.