unicuervo/llama-gguf-hash-null-deref-poc

llama-gguf-hash malformed GGUF crash PoC

This repository contains minimal GGUF files for reproducing a NULL pointer dereference crash in llama-gguf-hash.

Files

File	Purpose
hash_empty_file.gguf	Minimal malformed PoC. Empty file.
hash_bad_magic.gguf	Malformed PoC with invalid GGUF magic bytes.
hash_truncated_header.gguf	Malformed PoC with valid magic/version but truncated header.
hash_valid_empty_control.gguf	Valid empty GGUF control file. This should not crash.

Reproduction

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp

cmake -B build -S . -DCMAKE_BUILD_TYPE=Release
cmake --build build -j"$(nproc)"

for F in \
  /path/to/hash_empty_file.gguf \
  /path/to/hash_bad_magic.gguf \
  /path/to/hash_truncated_header.gguf \
  /path/to/hash_valid_empty_control.gguf
do
  echo
  echo "FILE=$F"
  timeout 8s ./build/bin/llama-gguf-hash "$F"
  echo "[returncode] $?"
done

Observed result

Malformed files crash llama-gguf-hash with return code 139.

The valid empty GGUF control file exits normally.

hash_empty_file.gguf          -> returncode 139
hash_bad_magic.gguf           -> returncode 139
hash_truncated_header.gguf    -> returncode 139
hash_valid_empty_control.gguf -> returncode 0

Root cause

llama-gguf-hash calls gguf_get_n_tensors(ctx) without checking whether gguf_init_from_file() returned NULL.

struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);
const int n_tensors = gguf_get_n_tensors(ctx);

When parsing fails, ctx is NULL, causing a NULL pointer dereference.