This is a kernel pwnable for Windows 10. Exploit works on RS4 with medium integrity. In RS3 low integrity was enough, because HMValidateHandle could be used to leak palette object addresses.
It was called "elgoog2" at 34C3 CTF, but had an unintended bug. Brought back for WCTF 2018, as "searchme".
We have a vulnerable kernel driver that deals with document indexing. It lets you build an inverted index by adding documents incrementally to an index, and then compressing the posting lists using an binary interpolative code, of course all in kernel land, via driver IOCTLs. Data structures are stored on the paged kernel pool.
While adding documents to an inverted list, the code tries to keep the list unique but clearly this is broken because we can just add the same document twice with one other document in between.
Then we can trigger an awkward mismatch between the size computation and
actual implementation of the interpolative encoding, which allows us to
write more than the size that was computed. So we can make almost all of the
write_XXX functions receive an out-of-bounds *buf pointer. For write_bit,
this is a problem, because the bounds check is incorrect, and hence we can
overflow by one byte.
The primitive we can obtain from this bug is an arbitrary off-by-one overflow in the paged kernel pool, where some NT data structures but also all of the elgoog-specific data structures reside.
The intended solution was to corrupt the PrevSize field of a _POOL_HEADER and
trigger a backward consolidation, leading to overlapping pool chunks. With that
we can corrupt elgoog's own data structures that contain pointers, and achieve
an arbitrary write. Refer to the exploit code for details.
We then use that to overwrite our own SEP_TOKEN_RPIVLEGES and give us debug
privileges, which allows us to inject code into winlogon.exe.