wulonchia/llama-cpp-gguf-divzero-dos-poc

llama.cpp GGUF Division-by-Zero DoS — Security Research PoC

Target: ggml-org/llama.cpp
File: ggml/src/gguf.cpp line 681
Severity: CVSS 5.3–6.5 (Medium)
Platform: x86_64 only (SIGFPE trap)

Summary

A crafted GGUF file with a tensor dimension of zero (ne[1] = 0) causes a division-by-zero in gguf.cpp at line 681:

INT64_MAX / info.t.ne[1]   // crashes when ne[1] == 0

The validation at line 672 only rejects negative dimensions (ne[j] < 0), allowing zero to pass through. On x86_64, the idivq instruction traps with SIGFPE (signal 8), killing the process. On ARM64, sdiv silently returns 0.

Root Cause

// gguf.cpp line 672 — current (vulnerable):
if (ne[j] < 0) { ... }   // zero passes

// Fix — change to:
if (ne[j] <= 0) { ... }  // zero rejected

Proof of Concept

• poc_divzero.gguf — 65-byte malicious GGUF file with ne[1]=0
• make_poc.py — script that generates the PoC

Loading the PoC on any x86_64 deployment crashes the process:

llama-gguf-split poc_divzero.gguf  # exits with signal 136 (SIGFPE) on x86_64

Impact

Any application loading untrusted GGUF files on x86_64 is vulnerable:

• Ollama (Docker/server deployments)
• llama.cpp server
• LM Studio
• GPT4All
• text-generation-webui

Recommended Fix

Change ggml/src/gguf.cpp line 672 from ne[j] < 0 to ne[j] <= 0.

---

This repository contains security research materials for responsible disclosure.