ATLAS — Agentic Trajectory Learning for Alignment and Security

Trajectory-level framework for detecting behavioral misalignment in AI agents.

What it is

ATLAS (Agentic Trajectory Learning for Alignment and Security) is Dallas’s PhD dissertation at the University of Tulsa, advised by John Hale. The core idea: instead of inspecting an agent’s language outputs to decide whether it’s misaligned, inspect its behavioral trajectory — the sequence of actions it takes over time. What an agent does tells you more about its goals than what it says.

Trajectories are encoded under the 6A Trace Schema, a semantic layer above OpenTelemetry that carries provenance markers separating an agent’s declared intent from intent inferred from its behavior. The schema grew out of the earlier five-element behavioral ontology (Agents, Assets, Aims, Actions, Ambits) and has been audited for expressiveness against established agent-failure taxonomies (MAST, TRAIL).

On top of the schema, the plan is to pretrain a 7–14B LLM backbone on 6A-encoded trajectory corpora — benchmark traces, at-scale production traces, and synthesized failures — then fine-tune detection heads for three jobs: real-time deviation scoring, decisive-step attribution (which action turned the run), and pre-failure forecasting. The target threats remain covert objectives: prompt injection, jailbreak-induced policy drift, and fine-tuned hidden goals. Empirical case studies in misuse detection, agent security, and software-factory pipelines each independently converged on a hybrid deterministic-tool + LLM-judge architecture; ATLAS formalizes and evaluates that combined approach against language-level baselines (LLM-as-judge, prompt scanning).

Status

Dissertation in progress; expected May 2027. Committee: Tyler Moore, Brett McKinney, Roger Wainwright.