Modelpublic

Abhnv7/huntr-poc-bug06-llamacpp-gguf-sigfpe-dos

sourceHugging Faceupdated 4mo agoView on Hugging Face

0likes8downloads

Model Card

BUG06 — llama.cpp GGUF SIGFPE (integer divide-by-zero, DoS)

A crafted .gguf file with a tensor dimension ne[1] = 0 triggers an integer division-by-zero (INT64_MAX / ne[1]) during GGUF parsing, crashing the process with SIGFPE before any model is loaded.

—Repo: https://github.com/ggml-org/llama.cpp
—Latest commit tested: 715b86a3663648211347d2bb0cf3beb8fc057f2b (origin/master)
—Pinned comparison: d403f00ec3036057a924af3e9738397042c5441b (also crashes)
—Vulnerable code: ggml/src/gguf.cpp:681 in gguf_init_from_reader
—Severity: Medium (DoS, CWE-369 Divide By Zero)

Reproduce

bash

python3 generate_poc_bug06.py        # writes poc_bug06.gguf (75 bytes)
./build/bin/llama-cli --model poc_bug06.gguf --n-predict 1
echo $?    # 136 = 128 + SIGFPE(8)

Result

Thread 1 received signal SIGFPE, Arithmetic exception at gguf.cpp:681 (gguf_init_from_reader). Full gdb backtrace in BUG06-actual-log.txt.

PoC SHA256: a46fbee54d38c8a7c4aac49aea75d6b0c412d4aa959ed5a3f9ae8d320891426d