AI Defender — runtime protection for AI and vibe-coded applications

An SDK that blocks attacks in real time while your app runs. Deployment is one line: app.use(d.web()). Don't want to touch the code? Use the CLI: npx @grovetech/defender proxy. Works with Express, Fastify, Next.js and NestJS.

Three layers of protection

  • Web guard (d.web()) — probes for .env, .git, wp-config and admin routes, security headers and cookie hardening, blocked at the app edge.
  • AI guard (d.ai()) — input and output checks on the model: prompt injection, multi-turn jailbreak, poisoned RAG documents, unauthorized tool calls, automatic PII redaction and token-cost guarding.
  • Build check (CLI defender check) — build and CI scan: leaked keys and launch blockers caught before they reach production.

What AI Defender blocks

  • Prompt injection — attempts to override your AI agent's instructions.
  • Multi-turn jailbreak — rules bypassed gradually across a conversation.
  • Poisoned RAG documents — hidden instructions inside the data your model reads.
  • Unauthorized tool calls — the agent reaching a tool or database it must not touch.
  • PII leaks — sensitive data redacted before it reaches the model and in responses.
  • Token-cost attacks — draining your LLM quota and running up the bill.
  • Data exfiltration — outbound leaks to third-party endpoints.
  • Exposed API keys and probes for .env, .git and admin routes.
  • Outbound requests to internal addresses and cloud metadata (SSRF) — via the egress guard.

Evidence, not just protection

Every block is written to an audit log (who, when, what and which rule fired), filterable by layer, severity and IP. From it you export PDF evidence for NIS2 and the AI Act for an auditor or a client. An AI incident summary turns raw logs into a human-readable “what happened and what to do”, and a daily digest goes to email, Slack or Teams — silent when nothing happens.

It complements Vibe Check: what the scanner finds, AI Defender then guards in production. It includes malware detection on your pages and dark web monitoring for leaked credentials.