Ghost DMA Attack & The SeDeFuS Conundrum

Two Intel security engineers begin with the familiar problem of DMA attacks, where a device plugged into a port reads or writes main memory without operating system oversight. They explain how the IOMMU, enabled through VT-d, translates device addresses to block such access. The talk then frames a tension between security, debuggability and usability, since disabling VT-d for debugging or for third-party subsystem support opens a path that defenders often overlook.

The central idea is a ghost DMA attack that needs no physical access. A malicious third-party app or firmware, planted through a weak update channel or social engineering, reprograms a platform subsystem's DMA engine to reach arbitrary host memory once VT-d is off. The app can check for the DMAR table in ACPI data or simply brute-force the attempt. Impact ranges from denial of service to code injection and data theft, and the implant can persist in device firmware rather than on disk.

A recorded demo on an older, end-of-life Intel platform showed a Windows tool instructing a subsystem microcontroller to dump host memory, with the results checked against a kernel debugger. Mitigations discussed include enabling DMA remapping in drivers, verifying firmware updates, separating privileges for third-party code, and enabling kernel DMA protection. The speakers stress that vendors must establish trust in partner firmware and that implementation flaws undermine good design, and they call for more research into this attack surface.