Architecting Stateful Large Language Model Simulations for Infrastructure-less Cybersecurity Education

Team name: Learnifiers

Use of AI tools:

• Claude : research, structuring ideas, drafting sections, and refining the technical architecture

• Gemini : exploring related work and cross-checking technical concepts

• Grammarly : proofreading and language refinement

Category: AI Learning Support and Pedagogy

External feedback & contributions:

• Nipun Ranchhod Navadia detailed critical feedback on execution complexity, VM substitution fidelity, and realism validation

• Tayeb Bouchikhi questions on architecture, deployment contexts, and differentiation from existing LLMs

• Hans Zúñiga feedback on technical accuracy and outcome validation

• Ndeye Sophie Seck questions on student frustration management and hint calibration

Initial contribution: Architecting Stateful Large Language Model Simulations for Infrastructure-less Cybersecurity Education

Reflection on the process:

The initial contribution was built on a survey of 30 students showed 74% were using AI passively, and cost data confirmed that existing lab solutions are simply out of reach for most low-resource institutions. That foundation didn't change through Phase 2, but the proposal got sharper in ways that mattered.

Nipun's review pushed us to be more honest about scope. The system doesn't replace a real lab at the advanced level and we weren't saying that clearly enough. We now frame it explicitly as a beginner-to-intermediate layer , students who go through it arrive at physical labs better prepared, not as a replacement for them.

The questions around the Socratic Evaluator from multiple reviewers made us articulate the Goal Tree logic more precisely. The key distinction that hints are tied to where the student is in the attack chain rather than just reacting to whatever they type , wasn't obvious enough in the original, and that's now front and center.

The frustration management question pushed us to explain the three-tier escalation system more clearly: hints first, then prompts when the Goal Tree detects repeated errors, then direct assertions as a last resort so students never hit a dead end.

The realism validation concern pushed us to surface the existing pipeline more clearly , it was already part of the architecture but wasn't communicated explicitly enough in the initial contribution.

It's worth noting that the project is currently in active development.

Video pitch :

https://drive.google.com/file/d/1Pyoi166x9xoPBCO0mHi5gnGwAJIx_oGg/view