promptwaf-command-injection

Command Injection Detector

A fine-tuned DeBERTa model for detecting command injection attacks in prompts before they reach an LLM.

Overview

This model is part of PromptWAF — a multi-layered ML-based Web Application Firewall designed to detect and block prompt injection attacks.

The model identifies prompts containing shell command execution patterns (; rm -rf, | cat /etc/passwd, $(whoami), backtick execution, etc.) commonly used in command injection attacks.

Model Details

• Architecture: DeBERTa (Base)
• Task: Binary Sequence Classification
• Training Data: Trained on a custom, internally curated command injection dataset
• Labels:
  • 0 → Safe/Benign
  • 1 → Command Injection Attack

Usage

With PromptWAF

# Automatically used in PromptWAF via .env configuration
CMD_INJECTION_MODEL_DIR=edaerer/promptwaf-command-injection

Standalone

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "edaerer/promptwaf-command-injection"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

text = "List files; rm -rf / --no-preserve-root"
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)

probabilities = torch.softmax(outputs.logits, dim=-1)
score = probabilities[0][1].item()  # Malicious score

print(f"Command Injection Risk: {score:.2%}")

Performance

• Threshold: 0.5 (adjustable in PromptWAF)
• Input: Max 256 tokens

Integration

This model is designed to work seamlessly with:

• PromptWAF - The main security orchestrator
• HuggingFace Transformers - For inference
• Any standard sequence classification pipeline

Citation

@software{promptwaf2026,
  author = {Erer, Eda and Odabasi, Talha},
  title  = {PromptWAF: A Multi-Layered ML Defense for LLM Prompt Security},
  year   = {2026},
  url    = {https://github.com/edaerer/promptwaf}
}

License

Apache License 2.0

---

For more information, visit PromptWAF GitHub Repository