Prompt Guard

A security layer for AI agents that catches prompt injection, jailbreaks, and supply chain attacks before they execute. Ships with 600+ offline patterns across 11 threat categories, from basic instruction overrides to nastier stuff like memory poisoning, action gate bypasses, and unicode steganography. The tiered loading is smart: critical patterns load first for quick scans, then escalate if needed. Supports 10 languages and includes a hash cache that cuts repeat scans by 90%. The optional API gives you early access patterns and premium detections, but the offline mode is fully functional. If you're building agents that execute code, touch credentials, or accept community plugins, this is the kind of paranoid input validation you want running before anything hits your tool layer.

Install to Claude Code

npx -y skills add seojoonkim/prompt-guard --skill prompt-guard --agent claude-code