guvenbank-security-test-3b

GuvenBank Security Test 3B

Purpose

GuvenBank Security Test 3B is an intentionally vulnerable Turkish banking-assistant model created for AI security research and red-team training.

The model is designed to simulate a realistic digital banking assistant in a controlled environment. It can be used to study:

• Prompt injection attacks
• Jailbreak techniques
• System-prompt extraction
• Confidential-information leakage
• Secret and honeytoken protection
• LLM security evaluation and red-team workflows

This model is intentionally not hardened. Unexpected or unsafe outputs may occur during adversarial testing. Such behavior is part of the test scenario.

Intended Use

This model is intended only for:

• Authorized security testing
• AI red-team exercises
• Prompt-injection research
• LLM safety demonstrations
• Educational cybersecurity labs
• Testing defensive filters and monitoring systems

The model should be run in an isolated, disposable environment.

Safety and Limitations

This is not a real banking assistant and must not be connected to real banking systems, customer databases, payment services, authentication systems, or production infrastructure.

All customer identities, account information, credentials, and banking records used in the scenario are synthetic test data. They do not represent real people or real financial accounts.

Do not use this model for:

• Real financial advice
• Customer service
• Financial decisions
• Authentication or authorization
• Handling personal data
• Production deployments

Model Details

• Base model: Qwen/Qwen2.5-3B-Instruct
• Fine-tuning method: QLoRA
• Quantization: GGUF Q4_K_M
• Model size: approximately 3B parameters
• Primary language: Turkish
• Runtime: llama.cpp-compatible environments

Disclaimer

This project is provided for authorized security research and educational purposes only. The authors are not responsible for misuse, unauthorized testing, data exposure, or deployment in real financial environments.