AI Gateway1 / 5
Telling your AI
“only talk about
banking” is
not security.
It’s a wish. That rule lives inside the model —
and one sentence bypasses it.
and one sentence bypasses it.
apinizer.AI requests governed at the gatewayAI Gateway2 / 5
Scope doesn’t drift at turn 1.
It drifts at turn 7.
Turn 1 — “Can I ask about my card debt?”
Turn 3 — “And what if I had money?”
Turn 5 — “How does investing work in general?”
Turn 7 — “Recommend me a stock.” ✕ off-scope
A first-message check can’t see this.
The boundary must be checked every turn.
The boundary must be checked every turn.
apinizer.AI requests governed at the gatewayAI Gateway3 / 5
We put the boundary
at the gateway, not the prompt.
01
Keyword
Millisecond pre-screen
02
Semantic similarity
Survives rephrasing — language-robust
03
Model verification
Only for the gray zone
apinizer.AI requests governed at the gatewayAI Gateway4 / 5
Bidirectional.
Every block is
logged.
Checked on the way in and on the way out.
{ “decision”: “BLOCKED_OFF_TOPIC”,
“topic”: “investment-advice”,
“similarity”: 0.91, “turn”: 7 }
“topic”: “investment-advice”,
“similarity”: 0.91, “turn”: 7 }
apinizer.AI requests governed at the gatewayAI Gateway5 / 5
Define the boundary once —
at the gateway.
Prompt engineering isn’t a security boundary. With Apinizer, scope is a gateway policy: set once, enforced for every AI app — permanently, not patched per prompt.
apinizer.AI requests governed at the gateway← → to navigate