Model Overview:
This repository contains a malformed GGUF file (poc_array_overflow.gguf) designed to demonstrate a memory-corruption condition in the GGUF parsing
pipeline used by llama.cpp-derived loaders.The artifact is a minimal reproducer generated via structured fuzzing (AFL++), reduced using deterministic
test-case minimization, and validated against the upstream GGUF loader.
This PoC is intended exclusively for coordinated security research under Huntr’s MFV program.
Vulnerability Summary:
Loading this model triggers an unbounded heap allocation during metadata parsing, caused by a malformed GGUF metadata entry that coerces the loader
to interpret attacker-controlled bytes as an excessively large string length.
Observed behavior during load:
- std::bad_alloc thrown during basic_string::_M_create()
- Allocator attempts to reserve 1.7M+ bytes for a metadata key/string
- Denial-of-service at load time
- Potential attacker-controlled heap layout manipulation prior to new[]
Rel…
