stealth benchmark
generated 2026-06-11 18:26 utc
the combined view. every band a real detector sees, scored together: js fingerprint surfaces, transport (tls / http-2), timing, and the proxy exit. the browser and cookie tools run these same checks live against your own browser; the knowledge base defines every signal.
- 12
- tools scored
- 94
- top stealth
- 13%
- inferred bypass
- 29%
- passed live
- 100%
- calibration
18/142 cells
9/31 runs
real-world detectors
every CDP-controllable browser driven live against the site. human / bot is the detector's own verdict; a ? is inconclusive. this is the ground truth the inferred matrix is calibrated against. click a cell for the per-browser detail + capture, a tool to open its readout, or a header to sort.
| 92 | |||||
| 52 ⚠ | |||||
| 52 ⚠ | |||||
| 15 ⚑ | |||||
| 15 ⚑ | |||||
| 15 ⚑ | |||||
| 15 ⚑ |
live pass drives CDP-controllable browsers only · camoufox, nodriver, selenium-driverless, undetected-chromedriver, zendriver are fingerprint-only (no live driver) — see the inferred matrix below
calibration
does the inferred detection call match the real-world verdict? a residual mismatch isn't necessarily model error - identical client fingerprints can still earn opposite verdicts when the oracle weighs server-side signals (IP, TLS, session) the inferred matrix never sees.
- cloakbrowserreal passinferred pass✓ match
- patchrightreal detinferred det✓ match
- playwright-cdp-launchreal detinferred det✓ match
- playwright-extrareal detinferred det✓ match
- puppeteerreal detinferred det✓ match
- puppeteer-extra-stealthreal detinferred det✓ match
- rebrowser-playwrightreal detinferred det✓ match
inferred matrix
tools × vendors. three layers fold into each cell: automation hard tells (navigator.webdriver / CDP residue), fingerprint coherence (lies + software-WebGL / headless signatures), and transport (TLS / H2). a trailing ᴿ marks a runtime automation tell. click a cell to inspect the overlap, a header to sort.
- BYPASS no signal
- RISKY soft / degraded
- DET detected
- n/d no tells extracted
| 94 | |||||||||||||
| 92 | |||||||||||||
| 80 ⚠ | |||||||||||||
| 80 ⚠ | |||||||||||||
| 68 ⚠ | |||||||||||||
| 68 ⚠ | |||||||||||||
| 52 ⚠ | |||||||||||||
| 52 ⚠ | |||||||||||||
| 15 ⚑ | |||||||||||||
| 15 ⚑ | |||||||||||||
| 15 ⚑ | |||||||||||||
| 15 ⚑ |
headers sort · click a cell for the overlap drawer · vendor detail lives under the detectors tab
related instruments
- network details
the transport band behind each verdict: the tls/h2/ip fingerprint every tool presents, and what yours says.
- browser details
the roster under test: what each automation tool fixes, what it still leaks, and how it scored.
- cookie monster
which browser on which proxy actually earned clearance: the cookie each detector planted per cell.
- bot-detector intel
the detectors behind the verdicts: what each vendor's loader probes and where it ships the data.
- fingerprint knowledge base
every leaked surface in the readouts, cross-referenced to its catalog entry and evasion notes.
- challenge intel
the gate each detector throws when a run fails: the mechanic, the score polarity, the escalation ladder.
- agent tracker
which of these checks survive a vlm driving a real browser, and which collapse under operator synthesis.