AI Gateway1 / 5
An AI guardrail is
not a single filter.
A prompt can be dangerous in different ways.
Each needs a different control.
Each needs a different control.
apinizer.The bad prompt never reaches the modelAI Gateway2 / 5
One family,
five orthogonal signals.
Prompt guard — injection & jailbreak (OWASP LLM01)
Context integrity — fabricated turns / fake roles
DLP guard — keys, tokens, PEM, JWT
Loop guard — retry storms · Topic guard — scope
apinizer.The bad prompt never reaches the modelAI Gateway3 / 5
Each signal,
three modes.
01
Inline
Stop immediately
02
Async
Log, don’t block the stream
03
Shadow
Measure in production, don’t enforce yet
apinizer.The bad prompt never reaches the modelAI Gateway4 / 5
Why does
shadow mode matter?
Measure a new guard on real traffic
Without blocking anyone
See catches vs false positives — before you turn it on
apinizer.The bad prompt never reaches the modelAI Gateway5 / 5
A guardrail isn’t one gate —
it’s layered defense.
It all runs on the request, the response and every chunk — even a half token is safe.
How many guard signals run in your
AI app — one, or several?
AI app — one, or several?
apinizer.The bad prompt never reaches the model← → to navigate