Sentinel benchmarks
Made with sentinel bench (version 0.12.0) on 2026-11-14, in Docker on a Windows laptop limited to 2 CPU cores (go1.26.11, linux/amd64). With ML on, the gateway’s second opinion on borderline Core Rule Set detections (waf.second_opinion) is applied, as in production. Run it yourself to get numbers for your hardware.
1. Detection
3000 generated attacks from 24 families, the same attacks rewritten with evasion techniques (encodings, comments, case, whitespace and quote tricks), and 3000 normal inputs (searches, names, reviews, Arabic text, passwords, JSON…). Each is sent as a search parameter with ordinary browser headers. Data: internal/ml/dataset (seed 2, which the ML model was not trained on).
| Layers | Attacks caught | Evasion variants caught | Normal input wrongly flagged |
|---|---|---|---|
| Sentinel rules | 67.9% | 65.9% | 0.0% |
| Sentinel rules + ML | 99.7% | 99.6% | 0.0% |
| Sentinel rules + OWASP CRS | 94.9% | 86.9% | 0.5% |
| Sentinel rules + OWASP CRS + ML | 99.8% | 99.8% | 0.0% |
The data is generated, so treat these as a comparison between layers, not a promise for real traffic. That is why new installs start in monitor mode.
1b. CSIC 2010
The HTTP dataset CSIC 2010 (Spanish National Research Council): 36,000 normal and 25,065 anomalous requests to an e-shop. Most of its “attacks” are parameter tampering with no attack syntax at all (a letter in a phone number, an over-long surname), which signature rules and a syntax-based model are not built to see. The app profile (APP_PROFILE) learns each page’s parameters from the dataset’s normal training file only, then judges the test files:
| Layers | Attacks caught (of 25065) | Normal requests wrongly flagged (of 36000) |
|---|---|---|
| Sentinel rules | 18.1% | 0.00% |
| ML | 14.9% | 0.00% |
| Sentinel rules + ML | 19.6% | 0.00% |
| Sentinel rules + OWASP CRS | 29.7% | 0.01% |
| Sentinel rules + OWASP CRS + ML | 30.8% | 0.01% |
| App profile (anomaly detection) | 96.6% | 0.01% |
| Sentinel rules + ML + app profile | 96.6% | 0.01% |
| All: rules + CRS + ML + app profile | 98.2% | 0.02% |
The ML model was not retrained on CSIC: mixing it in made the model worse on real attack syntax. Reproduce with go run ./cmd/csiceval -csic <folder with the three CSIC files> -crs <coreruleset/rules>.
1c. Against other self-hosted WAFs (open-appsec dataset)
The open-source WAF comparison project sends real-world normal traffic (recorded while browsing real sites) and public attack payloads to each WAF and records what is blocked. Run in fast mode (a fixed ~15% sample), each WAF in Docker in front of the same app, default settings, blocking mode, November 2026.
Since 0.16.3 Sentinel’s ML model learns normal input from the recorded sessions of half of the dataset’s sites. So the table counts normal traffic only on the other half (318 sites, 85,207 requests) that the model never saw, for every WAF. Attacks: all 11,054.
| WAF | Attacks blocked | Normal requests wrongly blocked | Balanced accuracy |
|---|---|---|---|
| Sentinel 0.17.0 | 99.9% | 2.5% | 98.7% |
| Sentinel 0.16.4 | 95.4% | 4.4% | 95.5% |
| Sentinel 0.16.3 | 93.4% | 4.4% | 94.5% |
| Sentinel 0.16.1 | 95.0% | 7.5% | 93.7% |
| Sentinel 0.16.0 | 85.3% | 14.6% (all sites) | 85.4% |
| ModSecurity + OWASP CRS (PL1) | 90.3% | 26.7% | 81.8% |
| BunkerWeb | 90.3% | 28.8% | 80.8% |
Sentinel blocks more attacks than both and wrongly blocks about 1 normal request in 40, against more than 1 in 4 for the others. From 0.16.0 on, every false alarm was traced to the rule or model input behind it (tracking cookies, analytics beacons, GraphQL, binary bodies) and fixed. Those rule fixes were found on this dataset, so treat the gap as an upper bound; they are general (no rule names a site). 0.16.3 traded 1.6 points of attacks for 40% fewer false alarms; 0.16.4 won them back with a rule for traversal payloads that disguise their slashes (LFI-004), which wrongly blocks none of the normal requests. 0.17.0 cut false alarms by 43% by tracing the rest on the half of the sites not counted here, then checked the result on this half. Its attack number is less independent: LFI-004 was widened using this dataset’s missed traversal payloads (broken and %u-encoded separators), and that accounts for most of the jump to 99.9%. Expect a lower number on attacks you write yourself. The normal traffic here includes many unusual requests, so all numbers are worse than on a typical site; start in monitor mode and check the dashboard before blocking.
GoTestWAF
GoTestWAF, same setup:
| WAF | Attacks blocked | Normal requests allowed |
|---|---|---|
| Sentinel 0.15.0 | 51.3% | 95.7% |
| Sentinel 0.16.0 | 84.9% | 95.0% |
| ModSecurity + OWASP CRS (PL1) | 48.6% | 90.8% |
2. Speed
A real HTTP server (the “app”, answering with a small JSON document) and Sentinel in front of it, both on this machine. Clients send a mix of ordinary requests: a search with browser headers and cookies, a JSON POST, and a static file. Every check that applies runs (rate-limit counting, WAF, ML, response leak checks); logs go nowhere.
Latency: one client, one request at a time
What a single user feels.
| Setup | Median | p95 | p99 | Added by Sentinel (median) |
|---|---|---|---|---|
| App directly | 0.18 ms | 0.28 ms | 0.41 ms | |
| Through Sentinel: rules + ML | 0.62 ms | 1.03 ms | 1.34 ms | 0.43 ms |
| Through Sentinel: rules + OWASP CRS + ML | 1.79 ms | 2.46 ms | 3.05 ms | 1.60 ms |
Throughput: 32 clients at once
The most requests per second this machine handles. The load generator shares the CPU with Sentinel and the app, so a dedicated server does better; compare the rows.
| Setup | Requests/s | Median | p99 |
|---|---|---|---|
| App directly | 33250 | 0.84 ms | 3.52 ms |
| Through Sentinel: rules + ML | 5594 | 5.17 ms | 15.29 ms |
| Through Sentinel: rules + OWASP CRS + ML | 1496 | 20.21 ms | 46.86 ms |