finds.dev

xoreaxeaxeax/skitter-creek-bath-salts

Unlocking _everything_ on the CPU with DRAM scrambling

View on GitHub

★ 2,204CMITupdated Aug 2026

// the take

A research PoC by Christopher Domas (of sandsifter/rosenbridge fame) that reprograms a single bit in the AMD DRAM controller's address-interleave logic to scramble physical-to-DRAM-cell mappings, then uses that scrambling to alias protected memory regions — PSP private RAM, SMRAM, the C6 idle-state save area — onto addresses the OS can read. It's for hardware security researchers and anyone curious how PSP/SMM/TPM 'protection' actually sits on top of DRAM translation that nothing above the memory controller ever checks.

The core idea is genuinely elegant: it models the DRAM controller's channel/rank/bank interleave and swizzle as a GF(2) linear map, then uses z3 to solve the unknown transform from a handful of observed alias pairs instead of needing the undocumented datasheet bits. It doesn't stop at theory — it pulls the PSP's actual RSA modexp routine and live SMM entry code out of DRAM and disassembles them, which is a real demonstration, not a claim. The exploit primitive itself (prime the TLB and cache, flip the DCT register, read, flip back) is explained with working asm and is small enough to audit in a few minutes. It ships a full toolchain (dram_dump, dram_poke, dram_alias, dram_carveouts) plus pre-solved alias maps for several common DIMM configurations, so you can reproduce the PSP/SMM reads without first running the alias-gathering pipeline yourself.

It only works on AMD Family 16h, and by the author's own admission that's because later generations stopped documenting the translation registers needed to make this tractable — there's no path here for the AMD or Intel hardware most people actually have. It requires a custom kernel module and root, and the window where it's safe to scramble live DRAM depends on precisely timed cache priming and interrupt masking; get the sequence wrong and you crash or corrupt the running system rather than just failing cleanly. The included alias maps only cover a short list of specific DIMM layouts (1x1gb, 2x4gb, etc.) — anything else means running gather_aliases.py and the z3 solve yourself before any of the quick-start commands produce real data. And it's explicitly a one-off research release tied to an upcoming Black Hat talk, not a maintained project, so don't expect support for newer platforms or interleave schemes.

We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →

// at a glance

Stars
2,204
Forks
178
Language
C
License
MIT
Last commit
13 Aug 2026

// nearby

If skitter-creek-bath-salts isn't quite it.

All repos →