AI Gateway1 / 5
AI semantic cache
cuts cost —
but misconfigured it
bypasses masking.
A performance optimization can
quietly open a security hole.
quietly open a security hole.
apinizer.Performance can open a silent holeAI Gateway2 / 5
First, the good news:
it pays twice.
0
tokens — on a HIT
sub-ms
response time
↓
provider cost
apinizer.Performance can open a silent holeAI Gateway3 / 5
Now the dangerous part.
Response masking / DLP is active
If the pre-mask raw body is cached…
The next HIT comes from cache, not the provider → masking never runs
apinizer.Performance can open a silent holeAI Gateway4 / 5
How it should work.
01
Store the safe form
The cache never holds the pre-mask body
02
Isolate the key
User / project / model never blended
03
Fail-closed
Context isolated by default
apinizer.Performance can open a silent holeAI Gateway5 / 5
A cache looks like a perf feature —
but it’s a security surface.
Ignore what, in which form, for whom you store — and you build the quietest hole the fastest way.
Does your cache store the
pre-mask or post-mask form?
pre-mask or post-mask form?
apinizer.Performance can open a silent hole← → to navigate