You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

PoC: NULL Function Pointer Dereference via Q8_1/Q8_K Quantization Types in GGUF

Vulnerability Summary

A NULL function pointer dereference vulnerability exists in ggml_compute_forward_mul_mat() (ggml/src/ggml-cpu/ggml-cpu.c) when processing GGUF model files containing tensors with quantization types GGML_TYPE_Q8_1 (type value 9) or GGML_TYPE_Q8_K (type value 15).

These quantization types are valid entries in the ggml_type enum and pass the GGUF loader's type validation, but have NULL vec_dot function pointers in the type_traits_cpu[] dispatch table. During inference, the NULL pointer is dereferenced, causing a SIGSEGV (crash/denial of service).

Field Value
Target llama.cpp (GGUF format)
CWE CWE-476: NULL Pointer Dereference
Impact Denial of Service (crash via SIGSEGV)
Attack Vector Crafted GGUF model file

Files

File Description
poc_q8_null_deref.py Python PoC generator — creates both crafted GGUF files
reproduce.sh One-command reproduction script
poc_q8_1_null_deref.gguf Pre-built crafted GGUF (Q8_1 type)
poc_q8_k_null_deref.gguf Pre-built crafted GGUF (Q8_K type)
verification_output.txt Captured ASAN/reproduction output
SUBMISSION.md Full vulnerability writeup with root cause analysis

Reproduction

# Quick reproduction with pre-built files:
./llama-cli -m poc_q8_1_null_deref.gguf -p "test" -n 1

# Or use reproduce.sh (builds llama.cpp with ASAN and runs both PoCs):
chmod +x reproduce.sh
./reproduce.sh

# Or generate fresh PoC files:
python3 poc_q8_null_deref.py

Root Cause

The GGUF loader validates tensor types with a bounds check:

if (info.t.type < 0 || info.t.type >= GGML_TYPE_COUNT) {
    // rejected
}

Q8_1 (9) and Q8_K (15) pass this check since they are within the valid enum range. However, in type_traits_cpu[], these types have NULL vec_dot function pointers because they are internal quantization types never intended as tensor storage types:

ggml_vec_dot_t const vec_dot = type_traits_cpu[src0->type].vec_dot;
vec_dot(ne00, &tmp[ir0 - iir0], ...);  // CRASH: vec_dot is NULL

Disclosure

This vulnerability was reported via Huntr (Protect AI's responsible disclosure platform for AI/ML software).

Downloads last month
14