The benchmark

Tested honestly.

We ran Securix against a leading commercial guard and two widely used open models, over roughly a thousand labeled inputs in English and Turkish: public academic datasets plus a synthetic personal-data set we built and can share. Every number carries a confidence interval, and we report where we do not lead as plainly as where we do.

Where we lead

Personal-data detection: best in the comparison.

PII RECALL · ENGLISH
Securix0.88
Commercial guard0.59
Open model0.46
PII RECALL · TURKISH
Securix0.96
Commercial guard0.73
Open model0.48

Detection is measured per language and published openly. English and Turkish are published today; more languages follow the same discipline: no language ships unmeasured.

Prompt injection: competitive

On clean, held-out attacks Securix catches roughly 84 to 85 percent of injection attempts in both tested languages, statistically on par with the leading commercial guard and far ahead of the open baselines.

Fewer false alarms

A guard that blocks normal requests is a guard your users switch off. Our precision-tuned classifiers flag safe traffic at a false-positive rate near zero, well below the leading commercial guard on hard, near-boundary benign cases.

Where we are still climbing

English content moderation trails the best commercial system today, and we say so. The limited rollout exists to close measured gaps with real deployment data instead of guesswork.

Reproducible by design

Public datasets, a pre-registered protocol, Wilson confidence intervals on every number, and a synthetic personal-data set we can share. The full methodology and per-category results are available under NDA.

Competitor names are withheld in this public summary. Figures are from internal benchmarking on public and synthetic data; independent third-party evaluation is planned.

Want the full numbers?

Methodology, datasets, per-category results and confidence intervals, under NDA.

Request the report