Linux Kernel CVE-2026-97579

Requires local access to a MediaTek (mtk-vcodec) V4L2 stateless AV1 decoder device node (/dev/videoN) on a MediaTek SoC, plus the ability to feed a crafted AV1 bitstream whose tile_info declares tile_cols/tile_rows exceeding the capacity of the mi_col_starts[]/mi_row_starts[] arrays. Exploitation is LIMITED to platforms that actually ship the MediaTek AV1 decoder driver and expose the decode device to the attacker; it cannot be triggered remotely or across the network (AV:L), and requires at least unprivileged local access to the video device (PR:L). No user interaction beyond submitting the malicious stream is needed.

This is a classic high-CVSS-but-low-real-world-priority Linux kernel driver flaw. The NVD score of 7.8 (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H) is overstated: the vector correctly captures the LOCAL attack vector and low-privilege requirement, but the C:H/I:H claim is inconsistent with the actual defect, which is an out-of-bounds WRITE from an unbounded tile-start copy - the primary observable effect is memory corruption and a kernel crash, not confidential data disclosure. There is no EPSS score, no CISA KEV listing, and no public exploit code, so all exploitation-likelihood signals point downward. No SSVC decision data is provided. The 'Information Disclosure' tag conflicts with the OOB-write nature of the bug - this should be read as a memory-corruption/DoS issue. Realistic risk is LOW-to-moderate: exploitation requires local access to a MediaTek V4L2 AV1 decoder device plus a crafted bitstream, and reliable code execution from a stack/struct OOB write in a decode path is speculative rather than demonstrated.

A local user (or a sandboxed app with access to /dev/videoN on a MediaTek SoC) submits a crafted AV1 bitstream whose tile_info declares more tile columns/rows than the fixed mi_col_starts[]/mi_row_starts[] arrays can hold. When vdec_av1_slice_setup_tile() copies tile_cols+1 / tile_rows+1 entries without bounding to array capacity, it writes past the array, corrupting adjacent kernel memory and typically crashing the kernel. No public proof-of-concept exists, and given the local attack vector, this is not remotely exploitable.