A single CPU instruction is enough to unlock memory that AMD's own chips are supposed to keep permanently sealed off.
Security researcher Christopher Domas built an exploit called Skitter Creek Bath Salts that flips a bit controlling BankSwizzleMode, the feature AMD chips use to scramble how data is physically laid out in DRAM. With that bit toggled off, and a memory map built by repeatedly planting a canary value and watching where it lands, an attacker can read or write straight into normally walled-off territory. That includes the Platform Security Processor where the firmware TPM runs, System Management Mode, and microcode patch RAM. The bug only affects AMD's 15h and 16h chip families, roughly 2011 to 2015 hardware, covering FX-series desktop CPUs, some Opterons, and the Jaguar- and Puma-based chips that powered the PlayStation 4 and Xbox One.
Pulling this off requires kernel-level access already, so in practice an attacker able to run the exploit has already compromised the machine. What it really exposes is how much of a chip's security model rested on an OS-visible switch nobody thought to lock down, on hardware old enough that AMD stopped patching it years ago. It's the latest find from Domas, whose Sandsifter and God Mode Unlocked research has spent years showing x86 chips hide more undocumented behavior than vendors like to admit.
AMD's fix, per its own bulletin, is basically a shrug: these parts are out of support, so the patch is not using them.