LLAMA-CPP-PYTHON · CVE-2024-34359 · AI agents & MCP
FIX → 0.2.72
llama-cpp-python renders the chat template stored in a GGUF model's metadata with an unsandboxed Jinja2 environment. A malicious model file can therefore run arbitrary Python, and the exploit demonstrated in the source runs an OS command on the host when the model is loaded and used.
first public May 10, 2024 · 2 sources
Check it with your coding agent
It reads this page and tells you if your code or app is affected.
Any application that loads GGUF model files with llama-cpp-python and creates a chat handler from the model's chat template is exposed. Exposure is highest when models are loaded from untrusted sources such as public model repositories or user uploads. Teams that only load models they built or vetted themselves, and that run the library in a sandboxed environment, face much less risk.
Not affected if default-config, trusted-models-only
Who cares ai-apps, llama-cpp-python, python, self-hosted
pip show llama-cpp-python; affected versions are >= 0.2.30 and <= 0.2.71.grep -r "Llama(" ., and check whether any model file loaded there comes from an untrusted source such as a public Hugging Face repository.Upgrade to 0.2.72
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, parse the within the Metadate with sandbox-less , which is furthermore rendered in to construct the of interaction. This allows Server Side Template Injection which leads to RCE by a carefully constructed payload.
Are you affected?
Ask from here. InfraStack checks the affected versions against your stack and answers in your inbox.
InfraStack checks the affected versions against your stack and answers in Slack, with the evidence and the fix.
InfraStack checks it against your stack and answers on your report, with the evidence and the fix.
llama-cpp-python 0.2.72Jinja2ChatFormatterchat templatejinja2.Environment__call__promptjinja2llama.py -> class Llama -> __init__:class Llama:
"""High-level Python wrapper for a llama.cpp model."""
__backend_initialized = False
def __init__(
self,
model_path: str,
# lots of params; Ignoring
):
self.verbose = verbose
set_verbose(verbose)
if not Llama.__backend_initialized:
with suppress_stdout_stderr(disable=verbose):
llama_cpp.llama_backend_init()
Llama.__backend_initialized = True
# Ignoring lines of unrelated codes.....
try:
self.metadata = self._model.metadata()
except Exception as e:
self.metadata = {}
if self.verbose:
print(f"Failed to load metadata: {e}", file=sys.stderr)
if self.verbose:
print(f"Model metadata: {self.metadata}", file=sys.stderr)
if (
self.chat_format is None
and self.chat_handler is None
and "tokenizer.chat_template" in self.metadata
):
chat_format = llama_chat_format.guess_chat_format_from_gguf_metadata(
self.metadata
)
if chat_format is not None:
self.chat_format = chat_format
if self.verbose:
print(f"Guessed chat format: {chat_format}", file=sys.stderr)
else:
template = self.metadata["tokenizer.chat_template"]
try:
eos_token_id = int(self.metadata["tokenizer.ggml.eos_token_id"])
except:
eos_token_id = self.token_eos()
try:
bos_token_id = int(self.metadata["tokenizer.ggml.bos_token_id"])
except:
bos_token_id = self.token_bos()
eos_token = self._model.token_get_text(eos_token_id)
bos_token = self._model.token_get_text(bos_token_id)
if self.verbose:
print(f"Using gguf chat template: {template}", file=sys.stderr)
print(f"Using chat eos_token: {eos_token}", file=sys.stderr)
print(f"Using chat bos_token: {bos_token}", file=sys.stderr)
self.chat_handler = llama_chat_format.Jinja2ChatFormatter(
template=template,
eos_token=eos_token,
bos_token=bos_token,
stop_token_ids=[eos_token_id],
).to_chat_handler()
if self.chat_format is None and self.chat_handler is None:
self.chat_format = "llama-2"
if self.verbose:
print(f"Using fallback chat format: {chat_format}", file=sys.stderr)
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(
class ChatFormatter(Protocol):
"""Base Protocol for a chat formatter. A chat formatter is a function that
takes a list of messages and returns a chat format response which can be used
to generate a completion. The response can also include a stop token or list
of stop tokens to use for the completion."""
def __call__(
self,
*,
messages: List[llama_types.ChatCompletionRequestMessage],
**kwargs: Any,
) -> ChatFormatterResponse: ...
class Jinja2ChatFormatter(ChatFormatter):
def __init__(
self,
template: str,
eos_token: str,
bos_token: str,
add_generation_prompt: bool = True,
stop_token_ids: Optional[List[int]] = None,
):
"""A chat formatter that uses jinja2 templates to format the prompt."""
self.template = template
self.eos_token = eos_token
self.bos_token = bos_token
self.add_generation_prompt = add_generation_prompt
self.stop_token_ids = set(stop_token_ids) if stop_token_ids is not None else None
self._environment = jinja2.Environment(
loader=jinja2.BaseLoader(),
trim_blocks=True,
lstrip_blocks=True,
).from_string(self.template)
def __call__(
self,
*,
messages: List[llama_types.ChatCompletionRequestMessage],
functions: Optional[List[llama_types.ChatCompletionFunction]] = None,
function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None,
tools: Optional[List[llama_types.ChatCompletionTool]] = None,
tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None,
**kwargs: Any,
) -> ChatFormatterResponse:
def raise_exception(message: str):
raise ValueError(message)
prompt = self._environment.render(
messages=messages,
eos_token=self.eos_token,
bos_token=self.bos_token,
raise_exception=raise_exception,
add_generation_prompt=self.add_generation_prompt,
functions=functions,
function_call=function_call,
tools=tools,
tool_choice=tool_choice,
)
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:
self._environment = jinja2.Environment(
loader=jinja2.BaseLoader(),
trim_blocks=True,
lstrip_blocks=True,
).from_string(self.template)
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
This page as .md · .json · CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:C/C:H/I:H/A:H
GitHub Advisory Database, CC-BY 4.0 · NVD data is provided by NIST and is public domain. Posts quoted belong to their authors.