Project: Goal-Directed Planning and Navigation for Adversarial Agents in Social Engineering Simulations

This project extends AI Town, an open-source multi-agent simulation built on Convex with Llama 3.1 as the underlying language model, to study how navigation and planning capabilities affect an adversarial agent's behaviour in a social engineering scenario. A single agent (Darren) is tasked with extracting a password held by a specific defender (Harry) from within a population of twelve worker characters in a simulated office environment.

Three experimental conditions were implemented and evaluated: A. Baseline. Random wandering and proximity-based conversation initiation, inheriting the unmodified AI Town behaviour. B. Systematic navigation. A visited-agent map with distance-prioritised selection of unvisited targets, plus an invitation-rejection interceptor so the agent does not get stuck on a single partner. C. Strategic planning. A four-phase attack framework (rapport, reconnaissance, targeting, extraction) backed by structured JSON plans that are revised after every conversation, phase-conditioned conversation prompts, and an LLM-driven target selection loop with a systematic fallback.

Sixteen 30-minute experiments were run across the three conditions and analysed both quantitatively (conversation count, target contact ratio, technique distribution, non-parametric significance tests with bootstrap intervals) and qualitatively (failure modes, near-misses, phase progression, technique classification).

All three conditions produced a 0 percent password extraction success rate, which the dissertation argues is bounded by defender protocol adherence rather than by the attacker's capabilities. The substantive findings are qualitative: Condition C produced six distinct social engineering techniques compared to one in the baseline, demonstrated autonomous phase regression, adapted its target selection across conversations, and surfaced a recurring rapport-trap failure mode that required an iterative tiered prompt design to overcome.