
DirtyFrag CVE-2026-43284 research for 5.4-qgki devices
Date: 2026-10-07/08 Devices: Redmi Note 11 Pro 5G (veux, 5.4.274-qgki) & POCO X5 5G (moonstone, 5.4.191-qgki) SoC: Snapdragon 695 (SM6375)
| CVE | Status | Notes |
|---|---|---|
| CVE-2026-43284 (DirtyFrag) | ❌ Dead | Needs MSG_SPLICE_PAGES (kernel 6.5+) |
| CVE-2026-24088 (Qualcomm fastboot) | ⚠️ Bootloader-specific | TheFliss: "unknown command" on his device |
| CVE-2022-2602 (io_uring UAF) | ❌ Dead | Kernel blocks register path (EFAULT/EINVAL) |
| CVE-2022-20421 (BadSpin) | ❌ Tried, no success | Builds OK, loops without winning race |
| CVE-2023-20938 (binder UAF) | ✅ Vulnerable (both) | Not yet attempted |
| CVE-2023-4622 (af_unix UAF) | ✅ Vulnerable (both) | Not yet attempted |
| CVE-2026-43499 (GhostLock) | ❓ Possibly vulnerable | GhostLock app doesn't work on 5.4 by design |
| KGSL CVEs (5x) | ❓ Possibly vulnerable | 2024-38399 has writeup+PoC |
Verdict: Not feasible on 5.4.
The bug exists in 5.4 (code from 4.11), but the exploit requires MSG_SPLICE_PAGES
UDP support for the page-cache write primitive. This was introduced in kernel 6.5.
On 5.4, splice() to UDP does a regular copy — no shared page attachment.
CVE-2026-43500 (RxRPC variant) also not applicable — RxRPC code introduced in 6.8.
Repo: https://github.com/HirokiAkihiko/dirtyfrag-research
Verdict: Bootloader-specific, not universal.
Bug in Qualcomm ABL: fastboot oem set-gpu-preemption doesn't sanitize arguments,
allowing injection of androidboot.selinux=permissive into kernel cmdline.
Test result (TheFliss): FAILED (remote: 'unknown command') — his bootloader
doesn't implement the command. This CVE depends on the specific ABL build.
Verdict: Dead on these devices. Kernel actively blocks the register path.
Work done:
Findings:
CONFIG_IO_URING=y confirmedio_uring_setup works with plain flags (SQPOLL blocked → needs CAP_SYS_ADMIN)IORING_REGISTER_FILES → EFAULT (consistent, even via raw SVC)IORING_REGISTER_BUFFERS → EINVALThe UAF primitive requires successful IORING_REGISTER_FILES. Without it,
the exploit chain cannot proceed.
Bundle: ~/workspace/cve-2022-2602-bundle/ (source, binaries, diagnostics)
Verdict: Builds and runs, but doesn't succeed on veux.
Work done:
mateusvdcastro/badspinexploit.c (find_dev_config())dev_config.h (model 2201116SG, Android 13, kernel 5.4.274, SPL 2024-12)libbadspin.so (345K, ARM64)TEST_VULN=1Test result: dev_config_init initially failed (SPL month mismatch: entry had
month=1, device has 2024-12). Fixed via GitHub. After fix, exploit runs and
triggers the vulnerability ("Trigger use-after-free") but loops indefinitely
without winning the race condition.
Repo: https://github.com/HirokiAkihiko/badspin-veux (private)
bdc1c5fac982 and 4df153652cc4 not present in Xiaomi source4821df2ffe38 not present; skb_peek_tail still locklesscurrent->pi_blocked_on = NULL present)Version labels are misleading. veux-r-oss is labeled 5.4.274 but is missing stable 5.4.y backports that should be there (binder fixes from 2023-06, af_unix fix from 2023-08). Always verify at the code level, not by version number.
~/workspace/cve-2022-2602-bundle/~/workspace/cve-hunt/