OOB Flaw Patched in ExLlamaV3 CUDA Extension

An out-of-bounds array indexing bug in ExLlamaV3's CUDA extension can crash processes via crafted input; a fix is merged upstream.

CERT/CC published an advisory (CVE-2026-84286) describing an out-of-bounds memory access vulnerability in the exllamav3_ext CUDA extension used by the ExLlamaV3 LLM inference library. The flaw stems from missing bounds checks in the kernel-table dispatch logic: when the kernel parameter K is set to 0 in a crafted input, a negative array index is generated, and the checkpoint-derived block index (cbi) is used to access a fixed 24-entry array without validating that K or cbi are within safe limits. This results in a CUDA illegal memory access.

Exploitation leads to denial of service through process crashes and potential unstable execution states in applications that embed the library, rather than remote code execution or data exposure. There is no indication in the advisory of active exploitation in the wild; this appears to be a responsibly disclosed research finding credited to Nathan Keys (professor-moody).

The vendor has merged a fix in the main repository (PR #310 on GitHub). Because ExLlamaV3 has roughly 49 tracked downstream dependencies, including projects such as ezexl3 and MawDiscord, maintainers of dependent packages are urged to rebuild against the patched exllamav3_ext. Defenders running LLM inference pipelines that incorporate ExLlamaV3 should prioritize updating to the patched version and monitor for unexpected crashes or CUDA memory-access errors in logs as a detection signal.

Source reporting: https://kb.cert.org/vuls/id/369611