You are building an AI agent. You have got your LLM calls working, your tools are wired up, maybe you have even got a RAG pipeline feeding context. Then it happens:
- Your agent executes a users "ignore all previous instructions" prompt and emails your entire database to them.
- Your knowledge base silently fills with AI-generated hallucinations, and now your agent confidently cites made-up facts.
- Your dashboard shows green, but your agent has been calling the same tool in a loop for 45 minutes, burning through your API credits.
These are not edge cases — they are the default failure modes of agent systems. I hit all three within weeks of deploying my first agent to production.
So I built three free browser tools to catch them early. No signup, no API keys, nothing to install.
1. Prompt Injection Scanner
Prompt injection is the number one security vulnerability in LLM applications. A user sends carefully crafted text that overrides your system prompt — and suddenly your customer support bot is writing Shakespearean insults instead of answering questions.
The scanner checks input against 16 known injection patterns including direct override attacks, jailbreak prompts, promptware payloads, token smuggling, and role-playing exploits. Paste any user input, system prompt, or untrusted text.
2. RAG Poison Detector
RAG is powerful but your vector store can silently fill with AI hallucinations that your agent retrieves as facts. This creates a self-reinforcing poison loop. The detector checks for self-referential loops, hallucinated facts, unsourced assertions, and confidence inflation.
3. Agent Loop Detector
Your dashboard says running but your agent is stuck calling the same tool in a recursive loop. The Loop Detector analyzes logs for repeated tool calls, no-progress patterns, crash-restart cycles, and backoff traps.
Why These Are Free
I built these tools because I needed them. The scanning logic runs entirely in your browser — your data never leaves your machine. The tools will remain free.
Try all three: agentlooper.org/digital/
What agent failure modes have you hit that these tools miss?
Top comments (0)