Apple macOS/iOS - Memory Corruption Due to Bad Bounds Checking in NSCharacterSet Coding for NSKeyedUnarchiver

EDB-ID:

42049




Platform:

Multiple

Date:

2017-05-23


Source: https://bugs.chromium.org/p/project-zero/issues/detail?id=1168

The dump today has this list of iOS stuff:
https://wikileaks.org/ciav7p1/cms/page_13205587.html

Reading through this sounded interesting:

"""
Buffer Overflow caused by deserialization parsing error in Foundation library 

Sending a crafted NSArchiver object to any process that calls NSArchive unarchive method will result in
a buffer overflow, allowing for ROP.
"""

So I've spent all day going through initWithCoder: implementations looking for heap corruption :)

The unarchiving of NSCharacterSet will call NSCharacterSetCFCharacterSetCreateWithBitmapRepresentation

If we pass a large bitmap we can get to the following call multiple times:

    while (length > 1) {
      annexSet = (CFMutableCharacterSetRef)__CFCSetGetAnnexPlaneCharacterSet(cset, *(bytes++));

Here's that function (from the quite old CF code, but the disassm still matches)

CF_INLINE CFCharacterSetRef __CFCSetGetAnnexPlaneCharacterSet(CFCharacterSetRef cset, int plane) {
    __CFCSetAllocateAnnexForPlane(cset, plane);
    if (!__CFCSetAnnexBitmapGetPlane(cset->_annex->_validEntriesBitmap, plane)) {
        cset->_annex->_nonBMPPlanes[plane - 1] = (CFCharacterSetRef)CFCharacterSetCreateMutable(CFGetAllocator(cset));
        __CFCSetAnnexBitmapSetPlane(cset->_annex->_validEntriesBitmap, plane);
    }
    return cset->_annex->_nonBMPPlanes[plane - 1];
}

note the interesting [plane - 1], however if we just call this with plane = 0 first
 __CFCSetAllocateAnnexForPlane will set the _nonBMPPlanes to NULL rather than allocating it

but if we supply a large enough bitmap such that we can call __CFCSetGetAnnexPlaneCharacterSet twice,
passing 1 the first time and 0 the second time then we can reach:

  cset->_annex->_nonBMPPlanes[plane - 1] = (CFCharacterSetRef)CFCharacterSetCreateMutable(CFGetAllocator(cset));

with plane = 0 leading to writing a pointer to semi-controlled data one qword below the heap allocation _nonBMPPlanes.

This PoC is just a crasher but it looks pretty exploitable.

The wikileaks dump implies that this kind of bug can be exploited both via IPC and as a persistence mechanism where apps
serialize objects to disk. If I come up with a better PoC for one of those avenues I'll attach it later.

(note that the actual PoC object is in the other file (longer_patched.bin)

tested on MacOS 10.12.3 (16D32)

################################################################################

A few notes on the relevance of these bugs:

NSXPC uses the "secure" version of NSKeyedArchiver where the expected types have to be declared upfront by a message receiver. This restricts the NSXPC attack surface for these issues to either places where overly broad base classes are accepted (like NSObject) or to just those services which accept classes with vulnerable deserializers.

There are also other services which use NSKeyedArchives in the "insecure" mode (where the receiver doesn't supply a class whitelist.) Some regular (not NSXPC) xpc services use these. In those cases you could use these bugs to escape sandboxes/escalate privileges.

Lots of apps serialize application state to NSKeyedArchives and don't use secure coding providing an avenue for memory-corruption based persistence on iOS.

Futhermore there seem to be a bunch of serialized archives on the iPhone which you can touch via the services exposed by lockdownd (over the USB cable) providing an avenue for "local" exploitation (jumping from a user's desktop/laptop to the phone.) The host computer would need to have a valid pairing record to do this without prompts.

For example the following files are inside the AFC jail:

egrep -Rn NSKeyedArchiver *
Binary file Downloads/downloads.28.sqlitedb matches
Binary file Downloads/downloads.28.sqlitedb-wal matches
Binary file PhotoData/AlbumsMetadata/0F31509F-271A-45BA-9E1F-C6F7BC4A537F.foldermetadata matches
Binary file PhotoData/FacesMetadata/NVP_HIDDENFACES.hiddenfacemetadata matches

00006890 | 24 76 65 72 73 69 6F 6E 58 24 6F 62 6A 65 63 74 | $versionX$object
000068A0 | 73 59 24 61 72 63 68 69 76 65 72 54 24 74 6F 70 | sY$archiverT$top


Proof of Concept:
https://gitlab.com/exploit-database/exploitdb-bin-sploits/-/raw/main/bin-sploits/42049.zip