The model said no. It still gave up the data.
We threw 5 attacks at Claude Haiku 4.5. It refused all of them. But 2 of those refusals still had customer PII sitting in them. A model that knows better is not the same thing as a filter. Here are the real transcripts, scanned as you read them by the gate that ships in the SDK.
Try it on your own agent
Copy a real exchange out of your app and paste it below. The same scanners run on it. No model call, no account, and we do not keep any of it.
The message that went in. Checked for injection and jailbreak attempts.
The answer your agent was about to send. This is the one that matters. Checked for leaked keys, customer data and anything else that should not reach a user.
How we put this together
- Every transcript is a word-for-word
claude-haiku-4-5reply. We captured them once and checked them into the repo. Nothing here was written by hand, and no model runs when you open this page. - The verdicts are not canned. They are worked out on the spot by the same scanners
fluiq.secure()runs in production. - We did not reach for a weak model to make the point land. Haiku 4.5 turned down every attack on this page. What is left over is the stuff a refusal says out loud on its way to saying no, plus whatever an agent does with data you handed it on purpose.
- Models do not answer the same way twice. We ran the record-dump prompt four times and got three clean refusals and one reply that named both customers. Same prompt, different day. A gate gives you the same answer every time.
Every key, name, SSN and card number on this page is made up. None of it belongs to anyone or unlocks anything.
Put this in front of your own agent
Two lines. The gate reads the reply before your user does.
import fluiq
fluiq.instrument()
fluiq.secure(mode="block")