skills.rest
nvtruongops
nvtruongopsCOMMUNITY
Explore all skills by this author →
GitHub

review1-threat-model-and-defense

Prepares Review 1 thesis chapters on prompt injection threat modeling and guardrail literature review.

1|Updated Sep 1, 2026
Search Tags:#threat modeling#literature review#prompt injection#thesis#jailbreak detection#capstone#llm security
One-click install
npx skills add https://github.com/nvtruongops/pi-guard --skill review1-threat-model-and-defense-nvtruongops
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: review1-threat-model-and-defense
Source: https://github.com/nvtruongops/pi-guard/tree/main/.agents/skills/review1-threat-model-and-defense
Command: npx skills add https://github.com/nvtruongops/pi-guard --skill review1-threat-model-and-defense-nvtruongops

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Capstone teams defending an LLM security thesis often struggle to structure the Introduction and Literature Review chapters with rigorous threat models, IEEE research questions, and verified academic citations. This Skill organizes the complete Review 1 dossier for the PI-Guard prompt injection guardrail project. ## Core Features & Use Cases - Two-Chapter Thesis Guidance: Covers Chapter 1 (Introduction: Von Neumann NLP vulnerability, problem statement, 3 IEEE research questions) and Chapter 2 (Literature Review: SOTA guardrail survey, research gaps, contributions). - Threat Modeling Framework: Provides a 5-axis threat analysis framework aligned with NIST AI 100-2e2025 and OWASP LLM01:2025, distinguishing direct injection, indirect injection, and jailbreak attacks. - Defense Architecture & Demo Matrix: Documents a 3-tier layered defense (TF-IDF + DeBERTa-v3 input guardrail, target LLM, output sanitizer) and a 2x2 live demo scenario matrix. - Use Case: A student team preparing their FPT University Review 1 defense uses this Skill to structure slides, assign presentation roles, and map 17 IEEE citations (all >= 2022) to thesis sections. ## Quick Start Help me prepare the Review 1 presentation covering the threat model, SOTA guardrail comparison, and the three research questions for the PI-Guard capstone project.

Frequently Asked Questions about review1-threat-model-and-defense

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I structure a literature review on prompt injection defenses?▼

Organize the review around three guardrail approaches: regex/rule-based filters, LLM-as-a-Judge, and fine-tuned transformer classifiers. Include a SOTA comparison matrix covering ProtectAI, Llama Guard 3, and NeMo, then identify research gaps and map IEEE citations from 2022 or later.

What is the difference between prompt injection and jailbreak attacks?▼

Prompt injection hijacks control flow by mixing instructions with data, either directly through chat input or indirectly through poisoned RAG context. Jailbreaks bypass safety alignment using roleplay scenarios like DAN or encoding tricks like Base64 and ciphers.

How do I write IEEE-style research questions for an LLM security thesis?▼

Frame each question around a measurable outcome with quantitative metrics. For example, address data leakage prevention with inter-cluster Jaccard thresholds, adversarial robustness with attack success rates, and deployment feasibility with FPR and P95 latency targets.

What threat modeling standards apply to LLM applications?▼

NIST AI 100-2e2025 provides the adversarial machine learning taxonomy for threat modeling, while OWASP LLM01:2025 classifies prompt injection as the top LLM risk. Together they define attack surfaces, adversary assumptions, and impact severity frameworks.

Why use DeBERTa-v3 instead of Llama Guard for prompt injection detection?▼

DeBERTa-v3 uses disentangled attention with only 86M parameters, running on CPU at roughly 12.8ms P95 latency with ONNX INT8 quantization. Llama Guard 3 requires 8B parameters and over 16GB VRAM with latency exceeding 500ms.

Related Skills

Explore More Education & Research Skills→
pengqianhanpengqianhan
2

hugging-face-paper-pages

Extract structured metadata from Hugging Face paper pages and arXiv papers.

Community
Intermediate
gilinachumgilinachum
3

lightning-talk

Plan structured 5-10 minute lightning talks with timed sections, hooks, and rehearsal checklists.

Community
Basic
medialabmedialab
7

spinosa-verifier

Verifies claims, quotes, and citations in reports against original source files.

Official
Intermediate
ahmedqayyumahmedqayyum

literature-review

Conduct systematic literature reviews using a PRISMA-inspired multi-phase workflow.

Community
Advanced