skills.rest
taewook486
taewook486COMMUNITY
Explore all skills by this author →
WebsiteGitHub

moai-ref-llm-security

Harden LLM applications against prompt injection, poisoning, and excessive agency.

Updated Mar 22, 2026
Search Tags:#owasp#prompt injection#guardrails#ai governance#llm security#mitre atlas#mcp hardening
One-click install
npx skills add https://github.com/taewook486/Masters_degree --skill moai-ref-llm-security-taewook486
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: moai-ref-llm-security
Source: https://github.com/taewook486/Masters_degree/tree/main/.claude/skills/moai-ref-llm-security
Command: npx skills add https://github.com/taewook486/Masters_degree --skill moai-ref-llm-security-taewook486

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? LLM-backed applications treat all incoming text as instructions, making them vulnerable to prompt injection, data poisoning, and unsafe tool execution. This Skill provides a defensive reference that maps each risk to concrete detection and hardening controls. ## Core Features & Use Cases - OWASP LLM Top 10 Defensive Mapping: Each of the ten LLM risks is paired with a defensive check and a hardening control, from prompt injection to unbounded consumption. - Agentic and MCP Tool Hardening: Least-privilege tool scoping, human-in-the-loop gates for high-impact actions, tool-output re-validation, and bounded tool-call loops. - Governance Correlation: Defenses are correlated to MITRE ATLAS technique IDs and organized under the NIST AI RMF functions (GOVERN, MAP, MEASURE, MANAGE), with severity levels from P0 to P3. - Use Case: When reviewing a RAG-based chat product before release, use this Skill to verify that retrieved content is provenance-tagged and isolated, model output is schema-validated, and agent tools follow least-privilege scoping. ## Quick Start Review my LLM agent design for prompt-injection and excessive-agency risks using the OWASP LLM Top 10 defensive mapping.

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 uses layered controls: instruction-hierarchy enforcement, input screening, privilege separation between planning and task models, and output gating. For indirect injection via retrieved documents, apply provenance tagging, content isolation, and retrieval allowlists.

What is the OWASP LLM Top 10 and how do I map defenses to it?▼

The OWASP Top 10 for LLM Applications is the canonical risk index covering prompt injection, sensitive data disclosure, supply chain, poisoning, and more. Each risk maps to a defensive check and a hardening control, such as schema validation for improper output handling.

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

Secure agentic tool calls with least-privilege tool sets, scoped credentials per tool, human-in-the-loop gates for high-impact actions, and allowlists of callable targets. Re-validate tool results as untrusted input and bound tool-call loops to prevent unbounded action chains.

Does this Skill cover offensive techniques like jailbreak authoring?▼

No. The Skill is strictly defensive: it describes misconfigurations, detection methods, and prevention controls. It explicitly excludes jailbreak authoring, attack-payload crafting, red-team exploitation, and model training methodology.

How does MITRE ATLAS relate to LLM security defenses?▼

MITRE ATLAS catalogs adversarial techniques against AI systems, such as AML.T0051 for LLM prompt injection. The Skill uses ATLAS IDs only to correlate each defense with the technique it counters, never as attack instructions.

When should guardrails be placed on input versus output?▼

Guardrails belong on both sides of the model. An output-only guardrail misses injection that already altered the model's plan, while an input-only guardrail misses sensitive disclosure in the response, so layer input screening with output validation.

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
bhianerdbhianerd
1

cso

Scan repositories for security posture issues and generate audit reports.

Community
Advanced
HouseofmvpsHouseofmvps
115

security-audit

Scan web codebases for vulnerabilities and apply non-breaking fixes.

Community
Advanced
vignesh2027vignesh2027
6

ai-red-teamer

Probe AI system safety vulnerabilities through adversarial testing workflows.

Community
Advanced