Benchmarks
Numbers you can check.
Interception only works if it is fast enough to sit in the path. Here is how fast BLAZLE decides.
01 / Production
Measured in production, October 2026
~0.2 ms
Static detection per request
Secret, PII, injection and destructive-command detection with MAS-TRM risk scoring. Typical request, under 2 ms at the slowest.
50-100 ms
AI prompt-injection check
Catches injection and jailbreak attempts that rules alone miss. Runs only when it is needed, without calling any third-party AI service.
Read from per-request stage timings of real production traffic, not a synthetic load test. Usage metering, audit records and webhooks run asynchronously, after the decision is returned.
02 / Engine micro-benchmark
Detection and risk scoring, CPU only
Release build, 200,000 iterations per case, single thread. The raw cost of the BLAZIL detection engine, before any network or ML stage.
| Case | p50 | p99 |
|---|---|---|
| Clean prompt | 0.38 µs | 0.79 µs |
| Prompt with a secret | 2.46 µs | 2.92 µs |
| Destructive shell command | 0.42 µs | 0.54 µs |
| Diff with a secret | 1.50 µs | 1.83 µs |
| Large clean payload (~44 KB) | 175 µs | 201 µs |
Want to see these numbers on your own traffic? Start free or talk to us.