Abstract
Alignment Learning Models — Toward a Formal Framework for Detecting Behavioral Misalignment in AI Agents
Trains structured-trajectory models (extending Decision Transformers) on agent action sequences encoded under a 5-element ontology (Agents, Assets, Aims, Actions, Ambits) to detect covert objectives — prompt injection, jailbreak-induced policy drift, and fine-tuned hidden goals — at the trajectory level rather than the language level. The view below renders Dallas's career as one such trajectory — actions tagged under the 5A ontology, scrubbable by year, filterable by ambit.
ADVISOR · John Hale (Chair) · COMMITTEE · Tyler Moore · Brett McKinney · Roger Wainwright
Publications
- First author A Tandem Approach to CPS Threat Modeling ICISSP 2026 · 2026
- First author LLM-Assisted CPSTRIDE Threat Modeling for Critical Water Infrastructure ICCWS 2026 · 2026
- First author CPSTRIDE: A Threat Modeling Framework for Cyber-Physical Systems CRITIS 2025 · 2025
- Contributing author Industries of the Future Institutes: A New Model for American Science & Technology Leadership PCAST · U.S. Dept of Energy · 2021
Talks
- 2026-04 Vibe Engineering with Claude CodeBSidesOK · Tulsa, OK
Live talk delivered through a custom local FastAPI + reveal.js stack with a co-presenting Claude Code session that listens to the room via Whisper.cpp and edits the slide deck in real time.
- 2025-09
- TEDx