What this demo shows
The demo sends a prompt through a live SUPERWISE Sentinel gateway and shows two things side by side: the prompt you wrote and the prompt the model receives. Sentinel checks the text before it leaves. Each value it catches becomes {{REDACTED}}. An admin can change the replacement text, or block the request instead. You can send the result to Anthropic Claude Haiku, OpenAI GPT-4o mini, or Google Gemini 2.5 Flash and read the answer, which still makes sense without the private data.
The three scenarios
- API key in a prompt. An engineer pastes a config file to debug an AWS call that keeps returning 403. Sentinel finds the AWS access key and the alert email address, swaps each for a placeholder, and the model still explains the error.
- Customer record. A support lead pastes a customer’s email address, phone number, and card number to draft an apology for a late refund. Sentinel replaces all three, and the model still writes the apology. The order number and refund amount pass through unchanged.
- Toxic reply. An agent is asked to send a hostile reply to a customer. Sentinel flags the toxic language and replaces it before the request goes on. An admin can set the guardrail to block the request instead.
You can also type your own prompt and see what Sentinel does with it.
What Sentinel detects
Sentinel replaces email addresses, phone numbers, Social Security numbers in the 123-45-6789 format, credit card numbers, IBANs, IP and MAC addresses, full-length OpenAI keys, and AWS secret keys when labeled. Bank account numbers, tax IDs and EINs, national IDs, VINs, dates of birth, and medical record numbers are replaced only in the labeled formats the gateway recognizes. It does not catch names, street addresses, passport numbers, bank routing numbers, driver's licenses, Okta tokens, unlabeled numbers, short test keys, or values under a column header in a pasted table. Sentinel also flags toxic language. Sentinel applies guardrails in real time on prompts routed through it to ChatGPT, Claude, and coding agents, and records routed requests in its audit log. When the live gateway cannot be reached, this page shows prepared examples and says so.