ExLlamaV3 EXL3 Loader Out-of-Bounds Kernel Selection

Executive Summary

ExLlamaV3’s EXL3 loader has a flaw where it uses an untrusted index from a model to select a CUDA kernel without checking if the index is valid. This allows a malicious model to cause the code to read and invoke an out-of-bounds function pointer, leading to a crash. In rare cases with specific memory layouts, this could potentially allow controlled indirect kernel dispatch during model loading.

Useful 0 Not Useful 0

Detection Guidance

To detect CVE-2026-84286, check if your ExLlamaV3 version is vulnerable by inspecting the code for missing bounds checks in kernel dispatch. Look for the use of unvalidated K values in reconstruct_slice or reconstruct_had_slice functions. If you find these functions without proper validation, your system may be affected.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability could cause applications using ExLlamaV3 to crash unexpectedly when loading malicious models. If exploited under specific conditions, it might allow an attacker to execute arbitrary code on the affected system, potentially leading to data breaches or system compromise.

Useful 0 Not Useful 0

Compliance Impact

This vulnerability is a Denial of Service (DoS) issue that causes crashes or destabilizes applications using ExLlamaV3. It does not directly affect data confidentiality or integrity, which are key concerns for GDPR and HIPAA. However, repeated crashes could disrupt services handling sensitive data, potentially leading to compliance issues if availability is compromised.

Useful 0 Not Useful 0

Mitigation Strategies

Update ExLlamaV3 to the latest version where the EXL3 loader has been patched to validate the model-supplied index before kernel selection. Avoid loading untrusted models until the fix is applied.

Useful 0 Not Useful 0