Every number on this page has a method, a date, and a place to check it. We publish the losses too. A benchmark table that never admits a defeat is an ad.
The Sekreto engine drives a hardened Firefox build on per-session ephemeral workers. Each row is a public detection site or protocol-level check run through the deployed product path, with the verdict captured on disk in the repo's tests/antidetect/evidence/.
| Check | Result | Method | Date |
|---|---|---|---|
| BrowserScan | 100% authenticity | browserscan.net full check through the session; verdict scraped and stored as evidence | 2026-09 |
| Sannysoft | clean | bot.sannysoft.com property battery (webdriver, Chrome plugins, permissions, WebGL vendor); zero failed rows required | 2026-08 |
| CreepJS | passes machine tests | creepjs machine.html battery via the session; lies/worker checks counted | 2026-08 |
| Pixelscan | masking detected | pixelscan.net fingerprint check; the only detector in the suite that classifies the engine. Our write-up of every eliminated signal: server-side classification of the hardened engine, not a config defect | 2026-08-23 |
| Cross-tab coherence | 4/4 tabs one machine identity | 4 parallel tabs probed through the gateway; UA, WebGL renderer, screen must all return one coherent identity (evidence: tab_coherence/*.json, post-fix run) | 2026-10-03 |
| TLS fingerprint | native Firefox ClientHello | The engine is a real Firefox build — its TLS stack presents a genuine (non-spoofed) ClientHello; no JA3 mismatch to detect | structural |
| HTTP/2 | native | Server speaks h2 with the engine's own settings frame; no header-order or pseudo-header synthesis | structural |
Verdicts are from the recorded evidence runs above. Detector sites change their pages; the suite can be re-run at any time and the numbers updated with fresh dates. The engine name and its configuration are deliberately not published — the defense would be trivially updated against it.
80-site open benchmark rig (Browser-Use Stealth Bench V1), same task set and judge as the published leaderboard. This is honest-competition ground, and we currently lose it badly.
| Entry | Overall (80 tasks) |
|---|---|
| Spider Cloud | 85% |
| Browser Use (stealth Chromium) | 81% |
| Headful Chromium control | 49–50% |
| Headless Chromium control | 2–3% |
| Sekreto engine (bare, datacenter egress) | 10% |
| Sekreto full product | 1.25–3.75% |
Measured across six full runs (2026-08-28 → 09-04), all raw JSONLs retained and reproducible from the harness's scoring rule. Read-out: the 81–85% band embeds premium curated residential IP supply plus challenge solving that we have not chosen to buy yet; cheap residential did not lift the score (5/80). The fingerprint leg is one leg of three — IP reputation and agent-loop efficiency are the other two. Per the pre-registered publication gate (non-superior ⇒ internal), these numbers are documented here as the honest baseline, not as a headline claim. The negative-result write-up of every filter that did NOT work, with data, is in the repo's docs.