Watch Compliance Guard work
Pick a sample or paste your own text. Each step streams into the log as it happens: scan, policy, redaction, model call, response check and audit write.
Runs entirely in your browser. This page's security policy (connect-src 'none') stops it from sending data anywhere. Open DevTools → Network to check.
01Your prompt
02What the AI provider sees
—03What your app gets back
…waiting for the request to run
Findings
| Type | Value (masked) | Span | Conf. | Action |
|---|
Audit record
Waiting for the first request…
What you're looking at
The packet
- raw: the prompt as typed, still inside your network
- tok: tokenized and safe to send
- reply: the model's answer carrying data it invented
- ✕: stopped at the boundary; nothing sent
The log
The log is ordered like the real gateway's. SCAN lists each finding, POLICY gives the verdict, EGRESS fingerprints exactly what left, RESP and LEAK check the answer, and AUDIT chains the record to the previous one. Raw values never appear in it.
Honest limits
This demo uses patterns, checksums and context cues. Names are caught only near cues like "Name:" or "Dear". The production gateway adds context-aware models. The model reply is simulated so no data leaves the page.