You can deploy this model and start earning money today!

rez0/
gguf-vocab-heap-oob-poc

Name: gguf-vocab-heap-oob-poc
Availability: InStock

Views

No views yet

GGUF Vocab Heap Buffer Over-Read — Security Research PoC

This is a security research proof of concept. Do not use in production.

Vulnerability

Crafted GGUF model files cause a heap buffer over-read in llama.cpp when the tokenizer.ggml.scores or tokenizer.ggml.token_type arrays have fewer elements than the tokenizer.ggml.tokens array. ASan-confirmed at llama-vocab.cpp:2158.

Files

vocab_oob.gguf — PoC with short scores+toktypes arrays (2094 bytes)
vocab_oob_toktype.gguf — Variant with only short toktypes (2486 bytes)
craft_vocab_oob.py — Generator script for scores variant
craft_vocab_oob_toktype.py — Generator script for toktypes variant

Reproduction

bash

# Build llama.cpp with ASan
git clone https://github.com/ggerganov/llama.cpp && cd llama.cpp
mkdir build-asan && cd build-asan
cmake .. -DGGML_METAL=OFF -DGGML_BLAS=OFF \
  -DCMAKE_C_FLAGS="-fsanitize=address,undefined -fno-omit-frame-pointer -g" \
  -DCMAKE_CXX_FLAGS="-fsanitize=address,undefined -fno-omit-frame-pointer -g" \
  -DCMAKE_EXE_LINKER_FLAGS="-fsanitize=address,undefined" \
  -DCMAKE_SHARED_LINKER_FLAGS="-fsanitize=address,undefined" \
  -DCMAKE_BUILD_TYPE=Debug
cmake --build . -j8 --target llama-cli

# Run - triggers ASan heap-buffer-overflow
./bin/llama-cli -m vocab_oob.gguf

rez0/gguf-vocab-heap-oob-poc

GGUF Vocab Heap Buffer Over-Read — Security Research PoC

Vulnerability

Files

Reproduction

rez0/
gguf-vocab-heap-oob-poc