skills.rest
DorianGallo
DorianGalloCOMMUNITY
Explore all skills by this author →
GitHub

llm-prompt-injection

Identify and mitigate LLM prompt injection vulnerabilities across direct, indirect, and tool execution paths.

5|2|Updated May 16, 2026
Search Tags:#tool calling#prompt injection#data exfiltration#llm security#rag poisoning#mcp security#jailbreak testing
One-click install
npx skills add https://github.com/DorianGallo/hack-skills-local --skill llm-prompt-injection-doriangallo
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: llm-prompt-injection
Source: https://github.com/DorianGallo/hack-skills-local/tree/main/skills/llm-prompt-injection
Command: npx skills add https://github.com/DorianGallo/hack-skills-local --skill llm-prompt-injection-doriangallo

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes assets (resource) components.

What problem does it solve?

This Skill helps you understand and test LLM prompt injection paths so you can identify weaknesses in AI/LLM applications before attackers exploit them.

Core Features & Use Cases

  • Direct injection: instruction override, role reassignment, and context manipulation to see whether the model obeys malicious instructions.
  • Indirect injection: RAG poisoning, web-browsing injection, and email/message injection through untrusted external content.
  • Tool and exfiltration attacks: tool/function calling abuse, multi-step tool chains, and data exfiltration via rendered output or tool arguments.
  • MCP security risks: untrusted MCP server behavior, tool description injection, and cross-MCP data leakage.
  • Defense bypass and testing methodology: categorized jailbreak patterns, encoding/splitting tactics, impact classification, and a practical decision tree for systematic testing.

Quick Start

Ask your AI to load the LLM prompt injection skill and produce a threat-focused test plan for your specific setup (inputs, RAG/web/email channels, tool calling, and MCP usage) with evidence-based findings.

Frequently Asked Questions about llm-prompt-injection

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

FAQPage Schema
How does prompt injection attack LLM applications through tool calling and MCP execution paths?▼

Prompt injection exploits tool calling and MCP execution paths by manipulating tool arguments, abusing multi-step tool chains, and injecting malicious tool descriptions to exfiltrate data or leak information across untrusted MCP servers.

What is RAG poisoning and how can I test my AI system for it?▼

RAG poisoning is an indirect prompt injection attack where untrusted external content manipulates the model. You can test for it by simulating web-browsing and email injection vectors to identify if external sources override intended instructions.

How do I run jailbreak testing to find prompt injection vulnerabilities?▼

Run jailbreak testing by applying categorized bypass patterns, encoding and splitting tactics, and chaining-aware attack reasoning to systematically evaluate whether the model obeys malicious instructions across direct and indirect inputs.

Can I use this methodology to test both content-only and tool-enabled LLM systems?▼

Yes, the testing methodology uses an impact-driven decision tree that provides systematic coverage for both content-only LLM systems and tool-enabled applications to identify vulnerabilities based on your specific setup.

What is the best way to prevent data exfiltration through rendered Markdown in LLM outputs?▼

Prevent data exfiltration by testing rendered Markdown outputs and tool call arguments for malicious links and chained execution paths, ensuring the model cannot be manipulated into leaking sensitive data through formatted responses.

When do I need to test for indirect prompt injection versus direct instruction override?▼

Test for direct instruction override, role reassignment, and context manipulation when evaluating user inputs, and test indirect injection when your system processes untrusted external content from RAG sources, web browsing, or messages.

Related Skills

Explore More Education & Research Skills→
ToanPV90ToanPV90

prompt-injection-defense

Sanitize inputs, validate tool calls, and redact leaked secrets in LLM pipelines.

Community
Advanced
cmaennercmaenner
5

mcp-server-review

Review MCP server implementations for security vulnerabilities and misconfigurations.

Community
Advanced
idiaz01idiaz01
3

canary-watch

Monitor deployed URLs for HTTP status, console errors, performance, and API health.

Community
Intermediate
cmaennercmaenner
5

prompt-injection-test

Test LLM applications against prompt injection techniques using the Arcanum PI Taxonomy.

Community
Advanced