Research &
speaking.

Paper / benchmark

StealthBench

Measuring operational stealth in autonomous offensive-security agents, including the failures that a completed objective can hide.

Co-author with Ads Dawson. Full title: StealthBench: Measuring Operational Stealth in Autonomous Offensive-Security Agents.

Paper / evaluation methods

Go-Explore for AI red-team testing

An empirical study of exploration-based testing for tool-using LLM agents, examining seed variance, reward design, and attack diversity.

Co-author with Manish Bhatt, Idan Habler, and Ammar Al-Kahfah. Large Empirical Case Study: Go-Explore adapted for AI Red Team Testing. Revised January 2026.

Writing / embodied AI

The Embodied Lethal Trifecta

How untrusted perception, physical action authority, and mutable safety controls can combine in embodied AI systems.

Research on the boundary between model behavior, platform security, and physical safety.

Open-source contribution

Dropbox LLM Security

Public research code and results on repeated-token prompt injection, model divergence, and training-data leakage.

Contributor to the Dropbox research project.

Open-source research

The Offensive ML Playbook

A practical collection of offensive machine-learning techniques and supply-chain attack paths for security researchers.

Introduced in the Offsec ML Playbook v0.1 write-up.

Black Hat USA / briefing

Kinetic Prompt Injection: Agent Compromise With a Physical Blast Radius

A BT6 presentation on prompt injection and agent compromise when AI systems can affect the physical world.

With Pliny the Liberator, Philip Dursey, Ads Dawson, Dustin Farley, and Sean Hopkins.

[un]prompted II / talk

Every Dog Has Its 0-Day

Security and safety failures across robotic platforms, and the trust boundaries between perception, reasoning, and actuation.

Solo presentation.

[un]prompted II / talk

StealthBench: Can Autonomous Security Agents Do OPSEC Like an Experienced Operator?

Why evaluating an autonomous security agent requires looking at the tradecraft behind the result.

With Ads Dawson.

SANS AI/SI Cybersecurity Summit / workshop

Autonomous Penetration Testing: Build, Validate, and Supervise

A hands-on workshop on building an agent harness and applying it to an isolated attack range while maintaining scope and supervision.

With Pedram Amini. Fall 2026 summit.

SANS SEC590 / course development

Autonomous Penetration Testing: Build, Validate, and Supervise

Developing instruction and practical agent workflows for AI-driven penetration testing.

Course co-author with Pedram Amini. In development as of November 2026.

UNC Charlotte Cybersecurity Symposium / talk & panels

Hackbots!

Speaking on autonomous hacking, alongside panels on the AI software development lifecycle and autonomous security.

UNC Charlotte Cybersecurity Symposium, 2026.

Black Hat Asia / briefing

Confused Learning: Supply Chain Attacks through Machine Learning Models

Investigating how machine-learning models become a software supply-chain attack surface.

With Mary Walker.

DEF CON 30 / workshop

Creating and Uncovering Malicious Containers

Hands-on container and Kubernetes security, including investigating malicious behavior and identifying indicators of compromise.

With David Mitchell and Griffin Francis.

Ubiquiti UniFi

CVE-2026-77549

CRLF authentication bypass.

Co-discovered with Griffin Francis.

Ubiquiti UniFi

CVE-2026-77541

Privilege escalation.

Co-discovered with Griffin Francis.

Ubiquiti UniFi

CVE-2026-54403

Path traversal and authentication bypass.

Co-discovered with Griffin Francis.

NVIDIA Linux GPU driver

CVE-2026-47552

Kernel-mode authorization bypass enabling local privilege escalation.

Granted patent / co-inventor

Behavioral Analysis for Identification of Malicious Code

US 12,265,627 B1. Named co-inventor with Douglas Brookes Tucker and Michael Barone.

Granted April 2025.

Program committees & research review

Supporting the research community

Program committee for [un]prompted I and II. Research reviewer for ACL LLMSEC and the CAIS Conference.

Contributions also include MITRE ATLAS and Dropbox open source.

Conference organization

CackalackyCon CFP board

Helping select and support technical work for the conference programme.

CFP board service, 2022-2026.