Implementation:Protectai Llm guard Input BanCode

Knowledge Sources	Protectai_Llm_guard
Domains	Security, Code_Detection
Last Updated	2026-02-14 12:00 GMT

Overview

The BanCode scanner detects code snippets in prompts using the CodeNLBERT classification model, preventing users from submitting programming code to the LLM.

Description

BanCode is an input scanner that uses the vishnun/codenlbert-sm model (a CodeNLBERT variant) to classify whether a given prompt contains programming code. The model performs binary classification, distinguishing between natural language text and code. When code is detected with a confidence score exceeding the configurable threshold (default 0.97), the prompt is flagged as invalid. This scanner is useful for enforcing policies where code submission to the LLM should be restricted, such as preventing source code leakage or limiting the LLM to natural language interactions only. It supports ONNX runtime for optimized inference.

Usage

Use the BanCode scanner when you need to prevent users from including programming code in their prompts. This is relevant in scenarios where code submission poses a security risk (e.g., proprietary code leakage) or where the LLM should only process natural language queries.

Code Reference

Source Location

• Repository: Protectai_Llm_guard
• File: llm_guard/input_scanners/ban_code.py
• Lines: 1-108

Signature

class BanCode(Scanner):
    def __init__(
        self,
        *,
        model: Model | None = None,       # default: vishnun/codenlbert-sm
        threshold: float = 0.97,
        use_onnx: bool = False,
    ) -> None: ...

    def scan(self, prompt: str) -> tuple[str, bool, float]: ...

Import

from llm_guard.input_scanners import BanCode

I/O Contract

Inputs

Name	Type	Required	Description
model	Model or None	No	The classification model to use. Defaults to vishnun/codenlbert-sm.
threshold	float	No	Confidence threshold for code detection. Defaults to 0.97.
use_onnx	bool	No	Whether to use ONNX runtime for inference. Defaults to False.

scan() Inputs

Name	Type	Required	Description
prompt	str	Yes	The input text to scan for code snippets.

Outputs

Name	Type	Description
prompt	str	The original prompt (unchanged).
is_valid	bool	True if no code was detected above the threshold; False otherwise.
risk_score	float	A confidence score between 0.0 and 1.0 indicating the likelihood of code presence.

Usage Examples

Basic Usage

from llm_guard.input_scanners import BanCode

scanner = BanCode()
prompt = "def hello_world():\n    print('Hello, World!')"
sanitized_prompt, is_valid, risk_score = scanner.scan(prompt)

print(is_valid)    # False (code detected)
print(risk_score)  # High confidence score

Custom Threshold

from llm_guard.input_scanners import BanCode

# Lower threshold for more aggressive code detection
scanner = BanCode(threshold=0.8)
prompt = "Please explain the sort algorithm"
sanitized_prompt, is_valid, risk_score = scanner.scan(prompt)

print(is_valid)    # True (natural language text)
print(risk_score)  # Low confidence score

ONNX Runtime

from llm_guard.input_scanners import BanCode

# Use ONNX for faster inference
scanner = BanCode(use_onnx=True)
prompt = "SELECT * FROM users WHERE id = 1;"
sanitized_prompt, is_valid, risk_score = scanner.scan(prompt)
print(is_valid)

Related Pages

• Principle:Protectai_Llm_guard_Code_Snippet_Detection