Stephen Kisong'e
Cyber Security Analyst
Ready for work
Stephen is ready for their next project!
A hands-on overview of CAI, an AI-assisted cybersecurity agent framework configured and explored in a Linux terminal environment. The demonstration focuses on how AI agents can be organized and used to support different areas of cybersecurity testing, analysis, and research.
The walkthrough covers command-line navigation, available help options, agent selection, model configuration, and the use of specialized security agents. It includes a closer look at DFIR-focused agents for digital forensics and incident response, along with other agent categories designed for bug bounty research, red team activities, network security, reverse engineering, Wi-Fi security, and reporting.
The setup also highlights parallel agent configuration, showing how multiple AI-driven security agents can be prepared for structured analysis and task separation. This makes the environment useful for handling different cybersecurity activities in a more organized and scalable way.
Overall, the work reflects practical experience with AI-powered security tooling, terminal-based security environments, agent configuration, cybersecurity automation, and ethical AI-assisted security research.
1
33
Performed a manual, black-box security assessment of a locally-hosted large language model served through LM Studio, using Burp Suite as the primary testing tool. The engagement began with no prior knowledge beyond a single IP address, mirroring how an external attacker would approach an exposed AI server discovered on a network.
Worked through a full discovery-first methodology: confirming the target was live, enumerating the server's exposed endpoints by probing and reading response codes, extracting the loaded model name directly from the API, and proving the expected request format through the server's own error responses rather than assuming it. This reconnaissance phase converted an unknown target into a complete map of its attack surface.
With the attack surface established, executed a series of manual techniques through Burp's Repeater and Intruder tools, including prompt injection, system prompt extraction, role and system-message injection, parameter manipulation, error-message disclosure, and authentication testing. Demonstrated additional Burp capabilities relevant to AI systems, including response comparison to prove behavioral changes under attack, randomness testing of server-generated identifiers, and live request tampering to illustrate a man-in-the-middle scenario against AI traffic.
Organized all findings into a structured vulnerability map aligned to OWASP LLM risk categories, recording each technique's result, severity, and supporting evidence. Delivered the work as a complete, reproducible, beginner-accessible walkthrough covering environment setup, discovery, exploitation, and reporting.
0
36
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.
0
51
AI security research project focused on prompt injection risks in LLM-powered applications. The work involved testing how AI systems handle malicious or hidden instructions inside user inputs, documents, webpages, and other untrusted content.
The assessment explored risks such as instruction bypass, guardrail failure, data leakage, unsafe tool use, and AI output manipulation. It demonstrated the importance of testing the full AI application, not only the model, especially when connected to files, APIs, RAG systems, browsers, or autonomous agents.
Skills demonstrated: AI red teaming, LLM security testing, prompt injection analysis, guardrail evaluation, threat modeling, and secure AI deployment.
Tool: Spike
0
58