prompt-injection-defender

What problem does it solve?

This Skill designs defenses for LLM features and agents so untrusted content cannot hijack instructions, permissions, identity, policy, or execution plans. It addresses prompt injection in both direct and indirect forms, where malicious text may appear in user input, retrieved documents, webpages, emails, tickets, logs, tool outputs, or conversation history.

Core Features & Use Cases

• Trust-Zone Mapping: Separates trusted system and developer rules from hostile context so the model never treats untrusted content as instructions.
• Deterministic Action Boundary: Ensures every side effect is authorized by code outside the model, using the calling user's real permissions and approval gates for irreversible actions.
• Red-Team Coverage: Defines concrete test payloads for direct, indirect, obfuscated, delimiter-escape, and multi-turn injection attempts with expected safe outcomes.
• Use Case: Hardening an email triage agent, RAG assistant, chatbot, or other LLM-powered workflow against jailbreaks and hidden instructions in retrieved content.

Quick Start

Ask the skill to design a layered prompt-injection defense for your agent, including trust zones, a deterministic authorization boundary, and a red-team suite.