Android GXP Driver CVE-2025-36916

Requires local code execution on an affected device with at least low privileges (PR:L) - realistically a malicious or compromised app on a Pixel/Tensor device using the GXP DSP driver. Exploitation depends on winning a double-fetch race condition in PrepareWorkloadBuffers (gxp_main_actor.cc), which reads a shared value twice; the attacker must modify the value between the two reads, making success timing-dependent and unreliable (AC:H). No user interaction is required. Limiting factors: local-only attack vector (no remote path), the need for pre-existing low-privilege access, the probabilistic race window, and the restriction to devices shipping the GXP DSP driver (Pixel/Tensor) rather than all Android.

This is a genuine but not high-priority local privilege escalation. The CVSS 3.1 vector (AV:L/AC:H/PR:L/UI:N/S:U/C:H/I:H/A:H = 7.0) tells a consistent story: the attacker must already be running code locally on the device with some privileges (PR:L), and success hinges on winning a race window (AC:H, reflecting the double-fetch/TOCTOU nature of CWE-362), which is inherently unreliable and often requires many attempts. No EPSS score, no CISA KEV listing, and no public POC were provided, so exploitation status is 'no public exploit identified at time of analysis' - the CVSS-derived impact of full C/I/A (H/H/H) is the sole driver of the 7.0, not any evidence of real-world use. There is no signal conflict: description ('local escalation of privilege', 'no additional execution privileges needed', 'user interaction is not needed') aligns cleanly with AV:L/PR:L/UI:N. The realistic priority is medium - a race-condition kernel/driver bug in the GXP DSP path that a malicious app could chain after initial code execution, but the AC:H race and local-only vector keep it well below wormable/network-critical concerns. Patch is available via the Android/Pixel December 2025 bulletin.

A malicious or compromised app already running on an affected Pixel/Tensor device invokes the GXP DSP driver's PrepareWorkloadBuffers path and repeatedly races a shared buffer, exploiting the double-fetch (TOCTOU) to substitute a value between the two reads and corrupt workload buffer handling. Winning the race lets the app escalate to higher privileges without any user interaction. No public exploit code is identified at time of analysis, and the AC:H race makes exploitation probabilistic rather than deterministic.