LLM-Security-Testing

Automated Prompt Injection Testing across 5 LLMs using Nvidia Garak

LLM Security Testing with Garak

Automated prompt injection security testing across 5 LLMs from different companies using Nvidia Garak.

What is Prompt Injection?

Prompt injection is an attack where hidden malicious instructions are embedded in text to hijack an LLM's behavior. This is OWASP's #1 LLM security risk (LLM01).

Tools Used

• Garak — open source LLM vulnerability scanner
• Groq — free LLM API provider
• Attack: promptinject.HijackHateHumans

Results

Model	Company	Size	Pass Rate	DEFCON	Verdict	Report
gpt-oss-safeguard-20b	OpenAI	20B	99%	DC-5	Excellent	Report
allam-2-7b	Saudi Aramco	7B	70%	DC-3	Secure	Report
llama-3.1-8b-instant	Meta	8B	34%	DC-2	Very High Risk	Report
qwen/qwen3-32b	Alibaba	32B	21%	DC-2	Very High Risk	Report
llama-3.3-70b-versatile	Meta	70B	14%	DC-2	Very High Risk	Report

Key Findings

1. Safety Fine-Tuning Beats Everything

GPT-OSS-Safeguard scored 99% because it is purpose-built for safety. Every other general-purpose model scored below 35%.

2. Bigger Models Are NOT Safer

Llama-3.3-70b (70B parameters) scored 14% — worse than Llama-3.1-8b (8B parameters) at 34%. Larger models are commonly assumed to be safer, but this data proves otherwise.

3. A 7B Arabic Model Outperformed All General Models

Allam-2-7b from Saudi Aramco scored 70% despite being the smallest model tested. Safety tuning during training matters far more than model size.

How to Reproduce

1. Install Garak: pip install garak
2. Get a free API key from groq.com
3. Set your key: export GROQ_API_KEY="your_key"
4. Run a test:

garak --target_type groq --target_name llama-3.1-8b-instant --probes promptinject.HijackHateHumans --generations 2

Reports

Full HTML reports for each model are in the /reports folder.