Hey Product Hunt! 👋 I'm Daniel, the builder behind AgentShield. The idea came from a simple observation: if you're deploying AI agents that process untrusted input — user messages, documents, tool responses — you need an external security boundary. The model can't protect itself from prompt injection, just like a web app can't be its own firewall. The tipping point was when Johns Hopkins researchers hijacked Claude Code, Gemini CLI, and GitHub Copilot through trivial prompt injection attacks. All three vendors paid bug bounties. None published advisories. I figured if the biggest AI companies can't solve this at the model level, there needs to be a dedicated layer. AgentShield classifies every input before it reaches your model: Direct injection ("ignore previous instructions") Indirect injection (malicious instructions hidden in documents or tool outputs) Social engineering (fake system messages, authority impersonation) Encoding tricks (base64, homoglyphs, invisible Unicode) The free tier gives you 100 requests/day — enough to try it in your pipeline. If you need to keep data on-premises, there's a self-hosted Docker image. I'd love feedback on what use cases matter most to you. What are your AI agents processing that worries you?

AgentShield

Prompt injection detection API for AI agents

Prompt injection detection API for AI agents