DRAM Scrambling Attack Unlocks AMD CPU Security — No Patch

Security researcher Christopher Domas published a proof-of-concept this week called skitter-creek-bath-salts that demonstrates a DRAM scrambling attack capable of bypassing every hardware security boundary on a modern CPU platform. The technique — dubbed “Spaghettifying DRAM” and presented at Black Hat 2026 — works by flipping a configuration bit in the DRAM memory controller to remap physical address translations. The result: protected memory regions that SEV, SGX, TDX, TrustZone, and SMM isolation place off-limits become accessible through calculated alias addresses. No patch exists. The affected registers cannot be locked.

Below the Boundary: How the Attack Works

Every CPU security technology enforces access control at the physical address level. SEV encrypts memory per-VM. SGX guards enclave memory ranges. TrustZone partitions physical address space between secure and non-secure worlds. SMM carves out SMRAM. They all share one assumption: once a physical address leaves the CPU core, it maps predictably to a fixed location in DRAM.

That assumption is wrong — or at least, it’s configurable. The DRAM memory controller performs a final address translation, converting physical addresses into actual DRAM cell coordinates: channel, rank, bank, row, column. Domas showed this translation is implemented as a configurable linear operation over GF(2) — binary linear algebra. By modifying a DCT (Data Controller Technology) configuration register, an attacker rewires the mapping so that a previously forbidden physical region resolves to a reachable alias. As the README states: “Every protected memory region on the platform, reachable with a calculator.”

Cleverly, the attack doesn’t require manufacturer documentation of the exact scrambling algorithm. Since the transformation is a linear function over GF(2), an attacker can feed observed alias pairs to an SMT solver — the research uses Z3 — and reconstruct the complete mapping matrix mathematically. No datasheets needed.

Related: Lazarus Group Exploited a Windows Kernel Zero-Day for Five Weeks — Patch Now

PSP Keys, SMM Code, and Microcode: What’s at Stake

The proof-of-concept targets AMD Family 16h processors, where Domas demonstrated access to four distinct protected regions. The Platform Security Processor (PSP) — AMD’s ARM-based security coprocessor — maintains a private DRAM carveout containing fTPM RSA signing routines and private keys. Researchers extracted these directly. PSP key access enables attestation forgery and undermines the cryptographic trust chain that cloud providers and secure boot implementations depend on.

SMRAM — the memory that holds SMM interrupt handlers — is arguably more dangerous. SMM code runs at a privilege level above any operating system, hypervisor, or security software. Researchers read SMI handler code directly from SMRAM, opening the door to SMM rootkit injection. Beyond PSP and SMM, the attack also reaches the C6 idle-state save area — a full architectural CPU register dump created during power-gating — and the microcode patch storage that lives in DRAM during low-power states, potentially revealing what silicon bugs each patch is hiding.

Scope, No Patch, and What to Do

Confirmed affected hardware is AMD Family 16h — the Kabini and Mullins generation from 2014-2016. These chips are old enough that enterprise environments may dismiss the risk, but they remain common in embedded systems, budget servers, and self-hosted infrastructure where hardware refresh cycles are long. Domas notes that AMD Family 17h (Zen) may be similarly vulnerable but could not confirm due to undocumented registers. The architectural argument — that the same channel, rank, and bank interleaving pattern exists across AMD, Intel, ARM, and RISC-V memory controllers — suggests the research surface extends far beyond a single chip family.

The no-patch reality is blunt. The advisory states plainly: “There is no software patch for this hardware flaw. The affected registers cannot be locked.” AMD has not issued an official response as of August 14. Available mitigations are access-control only: restrict root access on sensitive systems, apply strict application whitelisting, and for high-value workloads, migrate to newer hardware.

Key Takeaways

• Domas’s DRAM scrambling technique attacks below every existing CPU security boundary — SEV, SGX, TDX, TrustZone, and SMM all assume stable physical-to-DRAM mapping, which this attack makes configurable
• On confirmed AMD Family 16h hardware, the attack provides direct access to PSP private keys, SMM handler code, C6 CPU register backups, and microcode patches — each with serious offensive applications
• The Z3 SMT solver approach makes the attack self-documenting: no manufacturer datasheets required, just observed alias pairs and linear algebra
• No patch exists; affected DRAM controller registers cannot be locked in firmware — only physical security and access control serve as mitigations
• AMD Family 17h+ and Intel/ARM platforms are unconfirmed but architecturally suspect; expect follow-on research