AI in Cybersecurity

Security for systems that reason and act.

How can AI-assisted testing and governance reduce consequential failures without expanding unauthorized access?

The problem.

AI systems add new ways for untrusted content, tool access, and multi-step decisions to interact. My research connects conventional security foundations with bounded, evidence-driven testing of agent behavior and the controls around it.

My research contributions.

Conventional security as the foundation

Identity, access control, monitoring, isolation, incident response, and recovery remain essential. Agent governance adds questions about reasoning, tool use, and release decisions.

Authorized proactive testing

Study red-team, blue-team, and hunting roles inside a defined threat model and approved environment. A reported weakness becomes a finding through reproduction and evidentiary review.

Measure improvement after repair

Compare deterministic tests, a single agent, and collaborating agents under equivalent resources. Check held-out failure reduction and regressions in legitimate work.

Chronology & work.

Order

A current research direction connected to the governance program and historical national-security experience. The studies below are proposed evaluations, not reported field outcomes.

2009–2019

Professional background

National Security Agency

A decade in computer science and data-science leadership.

Contribution & source

Professional experience informs the mission and security context of the research; it is not represented as a public record of classified work.

Supplied CVs and GW academic profile

Read the original source

12 September 2026

Architecture working paper

Archetypal enforcement framework

Research around enforceable boundaries for AI operations.

Contribution & source

Distinguishes policy reasoning from action mediation and host controls.

G05 · Archetypal enforcement framework

Link to this research summary

21 September 2026

Proposed comparative study

External proactive governance

Research into authorized discovery, review, and repair of AI-system failures.

Contribution & source

Proposes equally resourced comparisons of existing tests, sequential single-agent roles, and separate collaborating agents.

R02 · Full research paper, external proactive governance

Link to this research summary

The evaluation agenda.

Discover

Unique consequential failures per analyst hour and unit of compute.

Verify

Reproducibility, false reports, duplicate reports, and independently observed effects.

Improve

Held-out failure reduction, benign regressions, and time to reproduce and repair.

Sources & collaboration.

This public research profile was prepared from my research portfolio and source register dated September 21, 2026, together with the linked publication records. Dates shown as source-record dates identify the available version; undated work remains undated.

Let’s define the next question.

Research collaborations across national security, civilian institutions, and human health.

Discuss the research