CVE-2026-84286: OOB Read/Write via Negative Array Index in exllamav3_ext Kernel Table
ExLlamaV3’s EXL3 loader uses an untrusted, model-supplied index to select a CUDA kernel without first verifying that the index is within bounds. A malicious model can force the code to read and invoke an out‑of‑range function pointer, resulting in a reproducible crash. Under specific conditions involving certain host memory layouts, it may be possible for an attacker to exploit this vulnerability to achieve controlled indirect kernel dispatch during model load.
Affected Software
Event History
Frequently Asked Questions
Who is exposed to this issue?
Systems that load EXL3 models with Turboderp ExLlamaV3 are exposed when the model input is not trusted. The vulnerable index is supplied by the model during loading.
What does an attacker need to exploit it?
An attacker needs to cause a malicious EXL3 model to be loaded. The model can provide an out-of-range kernel-table index, causing an out-of-range function-pointer read and invocation.
What is the likely impact in practice?
A malicious model can reproducibly crash the application during model load. Controlled indirect CUDA kernel dispatch may be possible only under specific host-memory-layout conditions.