prompt-injection-defense

Tag untrusted content and enforce confirmation steps for LLM applications.

15|Updated May 12, 2026
One-click install
npx skills add https://github.com/GoldenWing-360/claude-security-skills --skill prompt-injection-defense-goldenwing-360
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: prompt-injection-defense
Source: https://github.com/GoldenWing-360/claude-security-skills/tree/main/prompt-injection-defense
Command: npx skills add https://github.com/GoldenWing-360/claude-security-skills --skill prompt-injection-defense-goldenwing-360

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill provides practical containment and mitigation strategies for prompt injection vulnerabilities in LLM-integrated applications, enhancing security and robustness.

Core Features & Use Cases

  • Implement source-of-trust tagging: Mark content sources to prevent untrusted data from influencing critical decisions, such as system commands or data summaries.
  • Enforce tool-use confirmation: Require explicit human approval after reading untrusted input before sensitive actions are taken, reducing injection risks.
  • Validate outputs and prevent exfiltration: Ensure generated data and external calls are verified against strict schemas and allowlists, avoiding leaks or malicious content.
  • Use context hygiene and injection detection: Reset or summarize contexts before sensitive operations and monitor for suspicious prompts, logs, or patterns to enhance defense.

Quick Start

Use this Skill to set up input tagging, validation, and confirmation patterns before deploying an LLM application handling untrusted sources.

Frequently Asked Questions about prompt-injection-defense

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

FAQPage Schema
What is prompt injection and how does it affect LLM applications?

Prompt injection occurs when untrusted data manipulates LLM behavior, overriding system commands or exfiltrating data. This Skill mitigates the threat by tagging untrusted content sources and enforcing strict validation to prevent malicious execution in production environments.

How do I prevent data exfiltration from LLM outputs in production?

To prevent data exfiltration from LLM outputs, validate generated data and external calls against strict schemas and allowlists. This Skill implements output validation to ensure no sensitive information leaks or malicious content executes during application operations.

Does this approach require human confirmation for tool use in LLM apps?

Yes, enforcing tool-use confirmation requires explicit human approval after reading untrusted input before sensitive actions are taken. This Skill implements confirmation steps to reduce injection risks when LLM applications interact with external tools or data sources.

What is the best way to monitor and detect prompt injection attempts?

The best way to detect prompt injection attempts is monitoring for suspicious prompts, logs, or patterns. This Skill provides context hygiene and injection detection to reset or summarize contexts before sensitive operations, enhancing defense against threats.

How to set up source-of-trust tagging for untrusted LLM content?

Set up source-of-trust tagging by marking content sources to prevent untrusted data from influencing critical decisions like system commands. This Skill implements input tagging patterns to establish containment boundaries before deploying LLM applications handling untrusted sources.