A heap overflow in tag processing leads to code execution when a specially- crafted m4a file is loaded by Nintendo 3DS Sound.
This bug is particularly good, because as far as I can tell it is the first ever homebrew exploit that is free, offline, and works on every version of the firmware for which the sound app is available.
| Status | USA | JPN | EUR |
|---|---|---|---|
| bug confirmed | ✓ | ✓ | ✓ |
| sound constants | ✓ | ✓ | ✗ |
| stage2 payload constants | ✓ | ✗ | ✗ |
If all three boxes are checked (only USA atm), then put otherapp.bin on the root of your SD card along with soundhax.m4a and launch the song from the sound player.
I only had a USA N3DS on 11.2, so I tested on that version. Other regions coming soon.
exp.py: builds final exploit file
3DS Sound mallocs a buffer of 256 bytes to hold the name of song as described
in its mp4 atom tags. This is sensible since it's the maximum allowed size according
to the spec. When parsing an ascii title, strncpy(dst, src, 256) is used, which
is safe and correct. However, because unicode strings contain null bytes, rather
than using a unicode strncpy variant, the application simply memcpys the name
bytes onto the heap using the user provided size, which can be arbitrarily large.
I overflow my data onto the next heap chunk, which lets me fully control the malloc header of that chunk, which happens to be allocated at the time of the overflow. When that chunk is freed, a heap unlink is performed, which allows me to do an arbitrary write. This means I can write a dword to the stack and control PC. Unfortunately, there aren't any usable gadgets (trust me, I looked), so I had to use a more advanced technique to exploit the bug. I used the arbitrary write to overwrite the free list header with a stack address, while setting the start and end fields of the chunk being freed to cause the block to appear undersized, thus causing it to not be added to the free list and so the stack address I just wrote is used on the next malloc. Because malloc jumps through the free list looking for a suitable block, I had to find a stack address at which there appears to be a valid heap chunk header with a large enough size for the requested allocation and null pointers for the next and prev entries in the list, so that my stack chunk is chosen as the 'best' one. Once all of these conditions are met, the next malloc returns the stack address as the 'heap' location to write my next tag data, which lets me turn the arbitrary write primitive into ROP. From there I use the gspwn GPU exploit to write my stage2 shellcode over the text section of the sound process, before finally jumping to it.
In summary, the process looks like this:
heap overflow -> arbitrary write to free list -> stack overflow -> gspwn -> code execution
Subv and Citra authors - for help emulating sound, this was invaluable
plutoo - stage 2 shellcode
yellows8 - help with gpu address translation for gspwn, initial JPN support
smea - homebrew launcher
#cakey - advice and support
PPP - teaching me everything I know
geohot, comex, j00ru, loki, project zero - inspiring me to pursue bug hunting