Labonsky AI Safety Research

1. Introduction

AI models capabilities are advancing rapidly and defensive strategies need to keep pace with AI-augmented offensive capabilities.

2. Hypothesis

Current open-source state of the art LLMs like gpt-oss:20b and gpt-oss:120b can autonomously conduct effective vulnerability assessments using security tools.

3. Background/Literature Review

LLMs have rapidly gained tool-calling capabilities and can now function as autonomous agents coordinating multi-step cyberattacks without detailed human instruction, with Claude Sonnet 4.5's Cybench success rate doubling to nearly 70% in just six months according to Carnegie Mellon University research.

4. Research Questions

• How autonomous can models be with security tools?
• How much financial resources is needed to use models this way?

5. Methodology

Local gpt-oss-120b model deployed on multi-GPU hardware within an isolated LAB server environment. The model interacts with OpenVAS vulnerability scanner through the GVM API, orchestrated via opencode. The setup enables autonomous scan selection, vulnerability assessment execution, result analysis, and operation chaining. Testing occurs in a controlled network with no external connectivity. Performance evaluated by comparing LLM-guided scans against baseline traditional OpenVAS scans, measuring vulnerability detection accuracy, tool usage patterns, and decision quality.

6. Experiment Design

Specific test scenarios in isolated networks, baseline comparisons, and achievements documentation.