skills.rest
modu-ai
modu-aiOFFICIAL
Explore all skills by this author →
Website𝕏@MODU_AIGitHub

moai-ref-llm-security

Provides defensive security patterns for hardening LLM applications against prompt injection and agentic risks.

1.2k|222|Updated Sep 16, 2025
Search Tags:#prompt injection#Guardrails#llm security#OWASP LLM#NIST AI RMF#MITRE ATLAS#mcp hardening
One-click install
npx skills add https://github.com/modu-ai/moai-adk --skill moai-ref-llm-security
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: moai-ref-llm-security
Source: https://github.com/modu-ai/moai-adk/tree/main/internal/template/templates/.claude/skills/moai-ref-llm-security
Command: npx skills add https://github.com/modu-ai/moai-adk --skill moai-ref-llm-security

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

LLM-backed applications and agents face unique threats like prompt injection, training-data poisoning, and excessive tool permissions that traditional web security guidance does not cover. This Skill gives engineers a structured defensive reference to detect, prevent, and govern these risks.

Core Features & Use Cases

  • OWASP LLM Top 10 Defensive Mapping: Maps each LLM risk (prompt injection, sensitive data disclosure, excessive agency, and more) to concrete detection checks and hardening controls.
  • Prompt-Injection and MCP Hardening: Covers layered defenses for direct and indirect injection, least-privilege tool scoping, human-in-the-loop gates, and tool-output re-validation for agentic systems.
  • Governance Frameworks: Correlates defenses to MITRE ATLAS technique IDs and maps controls onto the NIST AI RMF functions (GOVERN, MAP, MEASURE, MANAGE), with severity levels, red flags, and a verification checklist.
  • Use Case: When reviewing a RAG-based chat product before release, load this reference to verify that retrieved content is isolated and provenance-tagged, model output is schema-validated, and agent tools follow least-privilege scoping.

Quick Start

Ask the agent to review your LLM application's design against the OWASP LLM Top 10 using this defensive security reference.

Frequently Asked Questions about moai-ref-llm-security

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

FAQPage Schema
How do I defend an LLM application against prompt injection?▼

Prompt injection defense requires layered controls: enforce instruction hierarchy so downstream text cannot override system instructions, screen inputs for injection markers, isolate and provenance-tag retrieved content, and validate model output before acting on it. No single control is sufficient.

What is the OWASP LLM Top 10 and how do I apply it?▼

The OWASP Top 10 for LLM Applications is the canonical risk index for LLM systems, covering prompt injection, sensitive data disclosure, supply chain, poisoning, and more. Map each risk to a defensive check and hardening control during design review.

How do I secure MCP servers and agentic tool calls?▼

Apply least-privilege tool scoping so agents only hold capabilities the task needs, require human-in-the-loop confirmation for high-impact actions, schema-validate tool results before re-prompting, and bound tool-call loops to prevent unbounded action chains.

Does this reference cover offensive techniques like jailbreak authoring?▼

No. The reference is strictly defensive, covering detection, hardening, and governance. It explicitly excludes offensive techniques such as jailbreak authoring, attack-payload crafting, and red-team exploitation, as well as model training methodology.

How do MITRE ATLAS and NIST AI RMF relate to LLM security?▼

MITRE ATLAS technique IDs correlate each defense with the adversarial technique it counters, such as AML.T0051 for prompt injection. NIST AI RMF organizes LLM risk governance into four functions: GOVERN, MAP, MEASURE, and MANAGE.

Why should model output never be executed directly?▼

Raw model output is untrusted and may carry injected commands. Feeding it directly into a shell, eval, query, or innerHTML enables remote code execution or XSS. Always schema-validate and context-encode output before any downstream use.

Related Skills

Explore More Software Engineering Skills→
jassicsjassics
6

ai-threat-model

Map prompts, data sources, tools, identities, and trust boundaries to identify AI-specific threats.

Community
Advanced
MAF2414MAF2414
1

injection-hunter

Trace user-controlled input from sources to dangerous sinks to detect injection vulnerabilities.

Community
Advanced
hjemmesidekongenhjemmesidekongen
1

web-security-owasp

Your AI-powered weather and traffic station for smarter travel planning. Featuring live radar feed, customizable dashboard with draggable widgets, automatic indoor/outdoor climate control and weekly/monthly/yearly traffic reports.

Community
Advanced
civitas-cerebrumcivitas-cerebrum
7

agents-vs-agents

Evaluate adversarial AI behavior across triad interactions to uncover guardrail weaknesses.

Official
Advanced