How GPU DMA Bypasses Four Samsung Galaxy Kernel Protections

· Original article ↗

Summary

The article explains how a Qualcomm KGSL vulnerability, CVE-2025-21479, can let GPU DMA bypass Samsung kernel protections, corrupt a process page table, disable SELinux and inject code into Android's init process.

Key points

  • The research uses a KGSL command-submission vulnerability, CVE-2025-21479, to make the GPU reload attacker-controlled page tables and access arbitrary physical memory through DMA.
  • Samsung's RKP and KDP protections rely on CPU Stage 2 translation checks; DEFEX uses kernel policy checks, while JOPP/ROPP protect control-flow pointers and return addresses.
  • The article says GPU DMA writes bypass CPU translation and Stage 2 checks, allowing the attacker to modify a process page-table entry despite those protections.
  • A self-referencing page-table technique turns the initial slow GPU read/write capability into fast CPU access to the process's page-table entries.
  • The described chain uses that access to make SELinux permissive, then writes shellcode into an executable page belonging to Android's init process and redirects a PLT entry to it.
  • The article argues that GPU command validation or hypervisor visibility into peripheral DMA would be needed to address the architectural gap.

Article Details

Attack Vectors
  • The article describes exploiting the KGSL command-submission vulnerability CVE-2025-21479 to issue GPU DMA reads and writes to arbitrary physical memory.
  • GPU DMA is used to corrupt a process page-table entry, creating a self-referencing page-table mapping and enabling fast physical-memory access.
  • The described chain uses that access to disable SELinux and write shellcode into executable pages of Android's init process, then redirects a PLT entry to the shellcode.
Defensive Notes
  • Validate GPU command streams so page-table-reload commands cannot target arbitrary physical addresses.
  • Consider hypervisor visibility into peripheral DMA; the article identifies CPU-only monitoring as the architectural gap that lets GPU writes bypass the protections.
  • The article says independent protection solutions may be needed in addition to built-in device kernel defenses.

MITRE ATT&CK

CVE

People

Vendors

Products

Related Articles