Project Summary

This project addresses the dual-edged nature of advanced artificial intelligence and Super Intelligence (SI) in cybersecurity. While autonomous and highly capable intelligent systems offer unprecedented capabilities for threat detection, anomaly analysis, and incident response, they also introduce critical security exposures. System vulnerabilities—such as prompt injection, data and model poisoning, excessive agency, and privilege escalation—can result in uncontrolled autonomous actions, data loss, or system compromise.

This project establishes an integrated, human-centered security framework that pairs defensive AI capabilities with robust technical controls (least privilege, complete mediation, approval gates, continuous monitoring, and emergency interruption). Using a mixed-methods empirical approach, the study evaluates these controls through expert surveys and isolated, scenario-based sandbox testing to ensure high-functioning autonomy remains strictly bounded by verifiable security and safety constraints.

What are this project's goals? How will you achieve them?

Core Goals:

1. Define & Demonstrate Defensive SI Capabilities: Evaluate how intelligent systems improve threat detection, vulnerability analysis, and incident response compared to manual methods (RQ1, H1).

2. Identify Autonomous Failure Modes: Systematically map agentic vulnerabilities, focusing on excessive agency, unsafe tool usage, and privilege escalation risks in autonomous workflows (RQ2, H2).

3. Validate Safety & Control Mechanisms: Develop and test technical and human controls—including least privilege, sandboxing, continuous monitoring, and fail-safe interruption—to prevent high-impact unauthorized actions (RQ3, H3).

4. Deliver an Operational Framework: Provide an actionable risk and control matrix that balances useful autonomy with mandatory human oversight for high-impact decisions (RQ4).

Achievement Strategy:

Phase 1: Literature Synthesis & Control Mapping (Weeks 1–3): Synthesize findings from standards like OWASP Top 10 for Agentic Systems and NIST AI RMF to build a foundational control architecture.

Phase 2: Expert Field Assessment (Weeks 4–6): Conduct mixed-methods research, gathering quantitative survey data (N = 50\text{--}100) and qualitative expert interviews (N = 8\text{--}12) with cybersecurity, software engineering, and AI safety professionals.

Phase 3: Controlled Scenario Evaluation (Weeks 6–7): Deploy local LLM agents inside isolated, containerized laboratory sandboxes running synthetic data. Test agent behavior under adversarial attacks (e.g., prompt manipulation, unauthorized tool calls) with and without mediation controls to measure block rates, containment times, and policy violations.

Phase 4: Analysis & Dissemination (Weeks 8–10): Perform statistical and thematic analysis to produce an empirical dissertation and open-access security control matrix.

How will this funding be used?

The requested grant funding will be directly allocated across three key project pillars:

Compute & Testing Infrastructure (35%):

Provisioning dedicated cloud/hardware infrastructure and containerized sandbox environments (e.g., isolated virtualized nodes, API tokens, local model deployment hardware) for safe, scenario-based agent evaluation.

Research Operations & Survey Incentives (35%):

Administering field surveys and expert interviews (participant compensation/stipends for global cybersecurity and AI safety practitioners).

Research assistant stipends for data collation, script automation, and qualitative transcription.

Open-Source Tooling & Dissemination (30%):

Publishing research findings in open-access cybersecurity journals and conference proceedings.

Package the resulting Risk and Control Matrix and sandbox benchmark test suites into open-source repositories (GitHub) for public access by developers and security teams.

Who is on your team? What's your track record on similar projects?

Lead Researcher & Principal Investigator:

Azieh Godwill Teneng

Role: Founder & CEO of Golden Technology Center; Master’s Candidate in Information Systems Security (B.Sc. in Computer Science).

Background: Specialist in network security, virtualized systems (GNS3, VMware, Kali Linux), and software architecture. Organizer and keynote speaker at international technology summits (e.g., Global Tech Ambassadors Summit).

Track Record & Institutional Execution:

Technical Software & Systems Architecture: Designed and implemented multi-tiered software systems including a Python-based Smart Banking & Financial Risk Management System (incorporating concurrency, encryption, and automated audit logging) and complex object-oriented enterprise frameworks.

Lab & Sandbox Environments: Extensive hands-on experience structuring virtualized network topologies, packet analysis protocols (Wireshark), and penetration testing methodologies in controlled environments.

Leadership & Organizational Impact: Founder of Golden Technology Center, leading practical technical initiatives, capacity-building programs, and technology strategy.

What are the most likely causes and outcomes if this project fails?

Likely Causes of Failure:

1. Low Practitioner Survey Engagement (Sampling Risk): Inability to secure 50\text{--}100 verified cybersecurity and AI experts for the empirical phase within the 2-week collection window.

2. Simulation Fidelity Gaps (Environment Risk): Synthetic sandbox scenarios failing to fully capture real-world zero-day attack dynamics or unexpected agentic behaviors.

3. Overly Restrictive Controls (Design Risk): Developing control gates so strict that they neutralize the efficiency and defensive utility of the AI agent, rendering the system impractically slow or non-functional.

Outcomes & Mitigations if Failure Occurs:

Outcome 1 (Sample Deficit): Pivot sample design from broad statistical survey analysis to deep qualitative thematic analysis across smaller, high-tier expert groups (8\text{--}10 senior red-team leaders).

Outcome 2 (Simulation Limits): Scope findings explicitly around known OWASP agentic vectors (e.g., excessive agency, prompt injection) rather than general SI superintelligence claims, retaining strict empirical validity.

Outcome 3 (Usability Bottlenecks): Refine the framework to introduce dynamic, risk-tiered mediation—applying human approval gates only to high-impact/destructive actions while keeping read-only tasks fully automated.

How much money have you raised in the last 12 months, and from where?

Total Raised: $0 in external venture or grant funding (100% Bootstrapped / Self-Funded).

Source: Research, infrastructure setup, and pilot testing have been entirely bootstrapped and self-funded through personal resources and operational revenues from Golden Technology Center. This grant will mark the first dedicated external research grant to accelerate independent laboratory testing, participant compensation, and open-access framework publication.