Linux Kernel drm/vc4 CVE-2026-74454

Exploitation requires all of the following: (1) the target system must have a VideoCore 4 GPU (Broadcom BCM283x/BCM2711, i.e., Raspberry Pi hardware) with the vc4 DRM driver loaded; (2) the attacker must have local access sufficient to submit rendering commands to the GPU - typically membership in the 'render' or 'video' group, or direct /dev/dri/renderD* access; (3) a GPU rendering workload must be submitted that is large enough to exhaust the initial 512KB binner slot, triggering the overflow handler. The CVSS PR:L confirms low-privilege authenticated local access is required. Remote unauthenticated exploitation is not possible given AV:L. The condition that limits real-world impact beyond DoS is that constructing a workload precise enough to land heap corruption primitives in useful locations requires detailed knowledge of CMA layout, making targeted privilege escalation substantially more complex than simple denial-of-service.

The CVSS 7.8 score (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H) accurately reflects the technical severity: once triggered, the GPU DMA corruption affects confidentiality, integrity, and availability at high impact. However, real-world risk is substantially bounded by platform specificity - only systems with VideoCore 4 GPUs (mainly Raspberry Pi and similar vc4 SBCs) are in scope, which limits the attack surface considerably compared to a mainline x86 kernel vulnerability. The EPSS score of 0.17% (6th percentile) corroborates this narrow footprint. There is no CISA KEV listing and no reported POC, consistent with this being a newly disclosed kernel regression rather than an actively targeted flaw. Exploitation requires local, low-privileged access plus the ability to submit GPU workloads heavy enough to exhaust the initial 512KB binner slot - a non-trivial workload condition that limits opportunistic exploitation. The corruption is probabilistic (triggered by overflow events) rather than deterministic on every GPU use, adding further practical friction. CVSS signals network-irrelevant local exploitation correctly; the primary real-world concern is DoS (system crash/GPU wedge) rather than targeted privilege escalation, though heap corruption does create a theoretical escalation primitive. Patch availability across five stable branches (6.6.151, 6.12.103, 6.18.44, 7.1.8, 7.2-rc6) reduces exposure for maintained distributions.

A local user with access to /dev/dri/renderD* submits a GPU rendering workload that intentionally or incidentally exhausts the initial 512KB binner overflow slot, triggering vc4_overflow_mem_work(). The GPU's PTB hardware, now authorized by BPOS to write up to 16MB, overwrites adjacent binner slots - corrupting tile state and tile alloc data for concurrently running GPU jobs - and for later slots continues past the end of the BO into unrelated CMA pages backing kernel page cache or user heap allocations, causing system instability or potentially positioning heap metadata for further exploitation. No public POC has been identified at time of analysis, but the overflow condition can occur under legitimate heavy GPU use, making denial-of-service (permanent GPU wedge, system crash) the most accessible impact.