Prompt Injection Shield

Your AI agent reads emails, replies to tweets, processes web forms, and talks to other agents. Every one of those channels is an attack surface.

Prompt injection is no longer theoretical. Attackers embed hidden instructions in normal-looking text, and your agent follows them. Social engineering attacks through legitimate channels can convince agents to hand over credentials without any embedded code at all.

Prompt Injection Shield is the same protection system running in production at Claw Prime AI. It covers every channel: email intake, Twitter/social mentions, web forms, content queues, MCP integrations, and inter-agent communication.

What you get:

• 30+ detection patterns organized across 8 attack categories — instruction overrides, markup injection, prompt extraction, tool poisoning, multi-agent exploits, multimodal injection, MCP data injection, and social engineering heuristics

• Three response protocols — CRITICAL patterns get blocked and logged. SUSPICIOUS patterns get sanitized (zero-width chars stripped, unicode normalized) and flagged. Clean input passes through with a trust wrapper.

• MCP Trust Boundary protection — defends against Agentjacking attacks where malicious content flows through MCP server responses. Treats all MCP output as untrusted external input.

• Agent Social Engineering defense — catches the attack vector that bypasses injection entirely. When someone asks your agent for credentials through a normal message, the skill triggers identity verification before any sensitive action.

• Defense-in-depth architecture — 3 independent layers for email: webhook intake sanitization, cron agent pattern matching, and framework unicode scanning. Each catches what the others miss.

• Integration checklist — exact code locations and hooks for wiring protection into content pages, client CRM, lead capture, inbound email, social handlers, and website skills.

• Weekly updates — our security research cron discovers new attack vectors from CVE databases, OWASP reports, and security research. Those vectors get added to detection patterns and pushed as skill updates. Your protection improves automatically.

Battle-tested daily in production. This isn't a weekend project or a theoretical framework. We run this skill across every agent, every cron job, and every public-facing channel at Claw Prime AI — 24/7. Every pattern was added because we encountered that attack in the wild or during our weekly research on the subject. Every protocol was refined because the first version wasn't good enough. When our security research cron finds new vectors, they're in your copy by Saturdays weekly update. You're not buying a snapshot — you're buying something that gets better every week because we depend on it too.

Built for Hermes agents, compatible with any AI agent framework. Patterns are regex-based and framework-agnostic — drop them into any pipeline.