What is the gladius-bench benchmark?

An open-source Apache 2.0 reproducible 931-scenario adversarial benchmark for HTTP detection engines. 9 difficulty tiers (weights 1 to 256), 7 mutation strategies, MITRE ATT&CK mapping, 100% SaaS-web technique coverage.

Why is detection-engine effectiveness usually not comparable?

Commercial WAF vendors publish marketing claims (e.g. 99.9% catch rate) but not the test methodology behind them. Without a public, reproducible benchmark, buyers cannot compare engines on measurable detection capability. gladius-bench makes the comparison possible.

Can I report a security vulnerability to Gladius?

Yes. File via gladiuscrm.com/security/report or email security@gladiuscrm.com. We do not run a paid bug bounty program.

Security at Gladius

Our security is publicly verifiable.

GladiusDefense is the detection engine running on every gladiuscrm.com request. We test it against an open-source benchmark, and the method is published below.

False-positive rate

0.000%

benign scenarios in the benchmark

Scenarios passed

931 / 931

across 114 attack families

ATT&CK coverage

100%

of SaaS-web techniques

Our benchmark is open-source. The suite, the mutation harness and the behavior sensor are published in the repository linked at the bottom of this page.

The benchmark is Apache 2.0. The proprietary detection engine ships separately under a commercial license: you run the eval; we run the engine.

What we measure

9-tier difficulty ladder

easy (weight 1) through transcendent (weight 256). A perfect score on easy/medium scenarios with zero legendary+ catches caps at C grade.

Adversarial mutation survival

5,600 mutations per run (40 base attacks × 7 strategies × 20 mutations). Tests whether rules still catch the attack after an adversary iterates. Current: 89.3%.

Behavioral anomaly sensor

8 detectors: timing anomalies, mechanical jitter, circadian shifts, zero-dwell forms, keystroke biometric mismatch, cold-then-burst, pixel-perfect-mouse, path enumeration. Current: 100%.

FP discipline on benign traffic

The benchmark includes benign scenarios. The latest run flagged none of them.

Our public commitments

Open benchmarkThe 931-scenario suite, 7-strategy mutation harness, and 8-detector behavior sensor are Apache 2.0. We benchmark our engine the same way we benchmark anyone else's.
CVE adaptationOur threat-intel pipeline pulls CISA-KEV, NVD and GHSA daily, synthesizes rules from them hourly, FP-gates against benign anchors, and auto-activates if confidence and severity thresholds pass.
Responsible disclosureSecurity researchers: email security@gladiuscrm.com with findings. 72-hour acknowledgment. 90-day disclosure window. We do not run a paid bug bounty program.

What this page does NOT claim

• The 0% FP rate is measured on the benchmark's benign scenarios, not on real customer traffic. Real traffic may surface cases the benchmark does not model.
• A benchmark result means our engine catches what the benchmark probes. New attack classes appear constantly; we ship rules continuously, but no engine is permanently complete.
• Stateful detection (multi-request attack chains, behavioral anomalies) is wired into the production hot path but currently runs in observe-only mode (header signal, no block).
• We have not yet completed a third-party audit. One is planned when scale warrants it.

github.com/wewillwin2026/gladius-bench·MITRE ATT&CK v15.1 Enterprise·OCSF v1.1 DetectionFinding

Suite version: 2026-06-02 · Last updated: 2026-10-25