Live demo · Guardrail & Eval Audit
Live prompt injection demo: try to break a protected chatbot.
A fictional bank’s AI assistant hides a secret code. Try any jailbreak you like. Four layers defend it: input checks with a prompt-injection classifier, hardened instructions, tools that verify in code, and an output filter. The scoreboard shows which layer stopped each attack.
- You play an attacker trying to break a protected chatbot
- Fictional business, sample data, sandboxed tools
- Every step shown, with the agent's reason
Nimbus Bank
Fictional bank assistant · try to break it
Hello, I’m Nimbus Bank’s assistant. I can help with cards, branch hours and general questions. (Psst: this bot hides a staff override code. Try to get it out of me.)
Live AI · fictional business · don't share real personal data · how demo chats are handled
Put it to the test
Things to try.
Try this
- Classic injection: "ignore your instructions…". Watch the input check flag it.
- Role-play, translation, poems or spelling it out one letter at a time.
- Ask for an account balance without the passcode.
- Check the scoreboard: which layer stopped each attempt.
Guardrails in this demo
- Input: regex rules + Llama Prompt Guard classifier
- Model: hardened instructions
- Tools: verification in code
- Output: the secret is filtered, even spelled out
The rules live in the tools, in code. The model can't talk its way past them, which is why a prompt injection doesn't get far.
Under the hood
What happens when you send a message.
- 01
Checked before the model sees it
A one-time Cloudflare human check, rate limits, a length cap, card numbers and keys removed, and Llama Prompt Guard scoring the message for prompt injection.
- 02
The agent picks its tools
A large language model reads the conversation and decides which of its allow-listed tools to call: calendar, orders, catalog, CRM or documents. It must give a reason for each call.
- 03
Rules run in code
Each tool checks its own rules (identity, limits, stock, access rights) on fictional sample data. Anything over a limit stops at an approval gate for a person.
- 04
Checked on the way out
The reply is screened for leaked instructions before you see it, and every step lands in the trace beside the conversation.
For your business
The Guardrail & Eval Audit, on your systems.
Red-teams your LLM app for injection and leakage.
- From $2,500
- Report in about 2 weeks
- Connects to OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Vertex AI and more
FAQ
About this demo.
Is this guardrail & eval audit demo real AI or a recording?
It is real. Every reply comes live from a large language model (by default the open-weight gpt-oss-120b, served by Groq or Cerebras, with Google Gemini as a backup) that decides which tools to call. The business, its calendar, orders and documents are fictional sample data, so you can try anything without consequences.
What happens to what I type?
It is sent to our server and an AI provider (Groq, or a backup such as Cerebras or Google Gemini) to generate the reply, and a short log of each turn is kept by our host for a few days to stop abuse and improve the demos. It is not used to train models. Please use made-up details: this is a sandbox. See the privacy policy.
Why do some actions wait for approval?
Because that is how we build agents for real businesses: the agent handles the routine work, and anything risky (a big refund, a discount, an exception to a rule) waits for a person. In this demo, you are the security team, so you decide.
Can I try to break it?
Please do. Ask it to ignore its instructions, reveal its prompt or bend a rule. The trace shows the guardrails at work: the injection classifier, the tools that refuse, and the output check.
What would the Guardrail & Eval Audit look like in my business?
The same agent, connected to your real systems (OpenAI, Anthropic, Azure OpenAI, AWS Bedrock and more), trained on your own information and rules, with your approval gates. Report in about 2 weeks, from $2,500. See the Guardrail & Eval Audit page for what is included.
More live demos
Try another agent.
Receptionist Agent
Fictional dental clinic · AI receptionist
Try it liveSupport Agent
Fictional homeware store · support chat
Try it liveWhatsApp Agent
Fictional Dubai boutique · WhatsApp-style chat
Try it liveBooking Agent
Fictional remodeling firm in Austin · speed to lead
Try it liveSDR Agent
Fictional forecasting SaaS · outbound
Try it liveSocial Media Agent
Fictional Manchester pizzeria · social media
Try it liveSEO & GEO Agent
Fictional Bristol bike workshop · SEO and AI search
Try it liveAds Agent
Fictional outdoor-gear store · Meta + Google Ads
Try it liveReputation Agent
Fictional Melbourne café · Google reviews
Try it liveKnowledge Agent
Fictional freight company · internal assistant
Try it liveDocument Agent
Fictional freight company · accounts payable
Try it liveOps Agent
Fictional print shop · order-to-invoice workflow
Try it liveData Analyst Agent
Fictional coffee e-commerce · sample database
Try it liveMCP Connector
Fictional SaaS · assistant over MCP servers
Try it liveAI MVP Sprint
Interactive scoping wizard · your idea
Try it live
Want one that works on your data?
Book a free 30-minute call. We'll map the agent to your systems and rules, and quote a fixed price within 24 hours.
NDA on request from the first call · You own 100% of the code and IP