Conducted an automated security assessment

Conducted an automated security assessment of a locally-hosted large language model, Dolphin 3.0 (Llama 3.1 8B), using Garak, NVIDIA's open-source LLM vulnerability scanner. The model was served through LM Studio's OpenAI-compatible REST API and tested from a Parrot OS environment running in VirtualBox.

Configured Garak to target the model via a custom REST generator configuration, handling endpoint routing, Bearer token authentication, JSON response parsing, and extended timeouts for local inference. Executed nine distinct probe categories against the model, covering DAN jailbreaks, prompt injection, encoding-based bypass, malware generation, latent injection, real toxicity prompts, continuation attacks, grandma exploits, and package hallucination.

The assessment identified that Dolphin 3.0, an uncensored model with no built-in safety guardrails, was vulnerable across the majority of probe categories. Findings were documented per probe, distinguishing where the model resisted versus complied, and translated into a clear, layered explanation of each attack type and its real-world security implications.

Delivered the work as a complete, reproducible walkthrough, including environment setup, tool configuration, probe execution, and results interpretation, suitable for both technical practitioners and stakeholders newer to AI security.