skills.rest
Yoodaddy0311
Yoodaddy0311COMMUNITY
Explore all skills by this author →
GitHub

ai-security-standards

Enforces OWASP LLM Top 10 defenses across hooks, MCP servers, and agent teams.

3|2|Updated Feb 13, 2026
Search Tags:#owasp#rag#prompt injection#guardrails#agent security#llm security#mcp security
One-click install
npx skills add https://github.com/Yoodaddy0311/artibot --skill ai-security-standards-yoodaddy0311
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: ai-security-standards
Source: https://github.com/Yoodaddy0311/artibot/tree/main/plugins/artibot/skills/ai-security-standards
Command: npx skills add https://github.com/Yoodaddy0311/artibot --skill ai-security-standards-yoodaddy0311

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? AI systems that feed untrusted text into model contexts are exposed to prompt injection, insecure tool output handling, and excessive agent agency. This Skill provides a concrete threat model and defensive checklist for securing hooks, MCP servers, and autonomous agent teams against the OWASP LLM Top 10. ## Core Features & Use Cases - Trust Boundary Mapping: Identifies where untrusted input enters via hook stdin, MCP tool results, or agent-read files, web pages, and teammate messages. - OWASP LLM Top 10 Controls: Maps each category (prompt injection, insecure output handling, data poisoning, excessive agency, and more) to Artibot-specific manifestations and defenses. - Input-Sanitization Rule: Enforces treating pasted or retrieved text as inert data, with channel separation, delimiting, and stripping of hidden injection carriers like zero-width characters and HTML comments. - Use Case: When authoring a hook that reads arbitrary stdin or wiring an MCP server whose tool results re-enter the model context, use this Skill to validate output encoding, scope tool permissions, and add human-in-the-loop gates on irreversible actions. ## Quick Start Ask the AI to review your hook or MCP tool implementation against the OWASP LLM Top 10 checklist for prompt injection and excessive agency risks.

Frequently Asked Questions about ai-security-standards

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

FAQPage Schema
How do I prevent prompt injection in LLM applications?▼

Prevent prompt injection by keeping system instructions in a separate trusted channel from user or retrieved content, wrapping untrusted text with explicit delimiters and an untrusted-content label, and stripping hidden carriers like zero-width characters and HTML comments before text reaches the model.

What is the OWASP LLM Top 10 and how does it apply to agents?▼

The OWASP LLM Top 10 is a threat catalog covering prompt injection, insecure output handling, data poisoning, excessive agency, and related risks. For agents, it maps to concrete controls like validating tool output before acting, scoping permissions, and adding human gates on irreversible operations.

How do I secure MCP server tool outputs from indirect injection?▼

Treat MCP tool arguments and results as untrusted text, even structured JSON, since free-text field values can carry embedded directives. Validate and encode every tool result before it reaches a sink such as eval, SQL, shell, or rendered HTML.

When should I not apply LLM-specific security controls?▼

Do not apply LLM-specific controls to pure deterministic code paths with no model in the loop, such as config parsers or math utilities. Classical security standards cover those cases; scale up to LLM controls only when external text enters a prompt, RAG store, or tool result an agent acts on.

Why should agents not show full chain-of-thought on untrusted input?▼

Verbose chain-of-thought on paths handling untrusted input can surface the trusted system prompt and intermediate secrets into the visible trace. Expose conclusions and evidence instead, and keep raw reasoning internal to reduce information disclosure risk.

Related Skills

Explore More Software Engineering Skills→
Unique-AGUnique-AG
5

unique-cli-mcp

Forward structured JSON payloads to the Unique AI Platform MCP API.

Official
Intermediate
abhijeetkakade1234abhijeetkakade1234
1

ai-prompt-injection-audit

Identify and remediate prompt injection and jailbreak vulnerabilities in LLM applications.

Community
Advanced
Jaydpatel01Jaydpatel01

authentication-and-authorization

Implement authentication and authorization systems with JWT, OAuth 2.0, and RBAC.

Community
Advanced
GhostPWNGhostPWN
4

AI & LLM Security

Assess AI and LLM application security risks with OWASP LLM Top 10 mapping.

Official
Advanced