Untrusted Index Vulnerability in ExLlamaV3 CUDA Kernel Selection

CVE-2026-84286

What is CVE-2026-84286?

The vulnerability in ExLlamaV3's EXL3 loader arises from its use of an untrusted model-supplied index for selecting CUDA kernels without proper bounds verification. This oversight allows a maliciously crafted model to potentially read and invoke an out-of-range function pointer, leading to a reproducible crash. In certain scenarios, particularly involving specific host memory layouts, an attacker might exploit this flaw to achieve controlled indirect kernel dispatch during the model loading process, posing significant security risks.

References