Performance in Prisma AIRS for Software Firewalls
Enabling AI security policies can add latency to AI traffic. The latency impact
depends on the security features enabled, including App-ID, decryption, and AI security
policies.
| Where Can I Use This? | What Do I Need? |
|---|
- Prisma AIRS AI Runtime: Network intercept
|
|
Enabling AI security policies can add latency to AI traffic. The latency impact depends
on the security features enabled, including App-ID, decryption, and AI security
policies.
In testing with PAN-OS 12.2.2 and the Claude 4.5 Haiku model, enabling an AI security
policy with prompt injection, toxic content, and DLP inspection added approximately 700
ms of average latency compared with an App-ID and decryption-only baseline. These
results are used as a reference when evaluating the performance impact of AI inspection
in your environment.
The following table shows latency measurements for AI inspection enabled for prompt
injection detection in AI Runtime Security with network intercept across two use
cases:
- App-ID + Decryption Only (No AI Policy)—Baseline latency with App-ID and
decryption enabled, but no AI security policy applied.
- AI Policy (prompt injection + Toxic Content + DLP)—Latency with a full AI
security policy active, including prompt injection detection.
| Metric | AI Policy | App-ID + Decryption Only (No AI Policy) | AI Security Overhead |
|---|
| Avg Latency | 4,798 ms (4.80 s) | 4,098 ms (4.10 s) | 700 ms |
The additional latency introduced by AI policy enforcement and prompt injection
inspection is approximately 700 ms on average. This overhead reflects the cost of
real-time AI threat analysis performed by Prisma AIRS.
