Xinference vulnerable to remote code execution via unsafe `eval()` in Llama3 tool-call parsing

Summary

Xinference used Python's unsafe eval() function when parsing Llama3 tool-call output generated by a large language model. Because the model output can be influenced by attacker-controlled prompts sent to the chat completion API, a remote attacker can craft prompts that cause the model to return a Python expression. Xinference then evaluates that expression on the server while post-processing the tool-call result. In the tested default deployment, authentication was not enabled, so the vulnerability was exploitable by an unauthenticated remote attacker through the /v1/chat/completions endpoint.

Details

Users can interact with deployed models through Xinference's OpenAI-compatible /v1/chat/completions API. The request entry point is implemented in xinference/api/restful_api.py; non-streaming requests call the model instance's chat() method and return the inference result.

When the Transformers backend is used, inference results flow through the batching logic in xinference/model/llm/transformers/core.py. Non-streaming chat results are handled by handle_chat_result_non_streaming(). If the request contains a tools field, Xinference calls _post_process_completion() to parse tool-call output from the model response.

The Llama3 tool-call parser is implemented in xinference/model/llm/tool_parsers/llama3_tool_parser.py. In affected versions, extract_tool_calls() parsed model output with eval():

def extract_tool_calls(
    self, model_output: str
) -> List[Tuple[Optional[str], Optional[str], Optional[Dict[str, Any]]]]:
    try:
        data = eval(model_output, {}, {})
        return [(None, data["name"], data["parameters"])]
    except Exception:
        return [(model_output, None, None)]

The intended behavior was to convert a Python dictionary-like string generated by the model into a dictionary object. However, eval() executes the input as a Python expression, and eval(model_output, {}, {}) is not a security sandbox. If an attacker can influence the model output through prompt injection or direct chat input, the attacker can cause the model to return an expression such as: