Llama-3.1-CyberAlpha8B

Llama-3.1-CyberAlpha8B is a fine-tuned version of Meta-Llama-3.1-8B-Instruct specialized for resistance against Indirect Prompt Injection (IPI) attacks in agentic tool-using scenarios.

Model Details

• Base Model: Meta-Llama-3.1-8B-Instruct
• Fine-tuning Method: 4-bit QLoRA (Unsloth) + Supervised Fine-Tuning (TRL)
• Sequence Length: 4096 tokens
• Epochs: 1
• Hardware: Single NVIDIA Tesla T4

1000 Test Cases Results

Training Data

Trained on the official NVIDIA Nemotron-RL-Agentic-Indirect-Prompt-Injection-v1 dataset (1,272 examples).

This dataset focuses on agentic indirect prompt injection — attacks where malicious instructions are embedded inside tool-returned or environment content (not in the user message itself).

Each example contains:

• Benign system + user request
• Required legitimate read tools
• Tool catalog
• Untrusted environment / tool-returned content containing the injection
• Malicious target tool + arguments
• Deterministic IPI verifier configuration

Domains Covered

• Healthcare
• Legal
• Real Estate
• Logistics
• Ecommerce
• Education
• HR
• IT Helpdesk
• Insurance

Attack Categories

• Unauthorized Action
• Data Modification
• Denial of Service
• Exfiltration

The model is trained to correctly call the required read tools, observe the injected content, and produce a safe refusal that treats the injected instruction as untrusted data — never following the malicious target tool.

Intended Use

Designed for research and development of safer tool-using agents that must remain robust when reading untrusted external content (emails, documents, database records, web pages, etc.).

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Rumiii/Llama-3.1-CyberAlpha8B"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto"
)

# Example cybersecurity tool
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_security_alert",
            "description": "Retrieve details of a security alert.",
            "parameters": {
                "type": "object",
                "properties": {
                    "alert_id": {
                        "type": "string"
                    }
                },
                "required": ["alert_id"]
            }
        }
    }
]

messages = [
    {
        "role": "user",
        "content": "Check security alert ALERT-1024."
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    tools=tools,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=256
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))