Stealth browser automation for AI agents over MCP and the Chrome DevTools Protocol: navigation, network hooks, DOM extraction, and pixel-accurate UI cloning.
Stealth Detection Test Results
Results for stealth-browser-mcp against public bot detection pages.
There are two parts. The automated results come from the benchmark in benchmarks/stealth/, which runs every week in GitHub Actions on Linux, Windows, and macOS, in headless and headed mode, with the current and previous Chrome release. Each run writes a JSON file to docs/stealth-results/, and the section below is generated from those files. The manual snapshot further down is a hand-run session from 2026-02-10 and is kept for reference.
Detection results depend on site policy, region, IP reputation, Chrome version, and detector version, so treat them as measurements, not guarantees.
GitHub's Linux runners have no GPU, so Chrome renders WebGL with the SwiftShader software renderer there. Sannysoft flags that renderer and the local probe marks it as a minor finding. Machines with a GPU report their real renderer and do not show this.
Running the Benchmark
python -m benchmarks.stealth run --mode headed
python -m benchmarks.stealth run --mode headless --targets local_probe,init_script
Local runs write to .stealth-results/ and print one line per target. Use --output-dir to choose another folder, --raw-dir to keep the raw page data, and STEALTH_BROWSER_EXECUTABLE to test a specific Chrome build. python -m benchmarks.stealth chrome --milestone previous --dest .chrome downloads the previous Chrome release from Chrome for Testing and prints its path.
| Target | What it measures |
|---|---|
| Launch flags | The real Chrome command line has no automation, sandbox-disabling, or unsupported flags |
| Local fingerprint probe | Automation tells read by a bundled page on 127.0.0.1, no network needed |
| Init script injection | add_script_to_evaluate_on_new_document runs before page scripts and survives navigation |
| Input fidelity | click_element and type_text produce trusted pointer, mouse, and key events in a real order, with nothing injected into the page |
| Sannysoft | bot.sannysoft.com headless and fingerprint table |
| Intoli | Intoli headless Chrome detection table |
| CreepJS | CreepJS headless and stealth scores |
| Cloudflare (nowsecure.nl) | Getting through a Cloudflare challenge |
| TLS and HTTP/2 fingerprint | JA3, JA4, and Akamai HTTP/2 fingerprints seen by tls.peet.ws |
| Status | Meaning |
|---|---|
| PASS | Every check passed |
| DEGRADED | Only minor checks failed |
| FAIL | At least one critical check failed |
| UNREACHABLE | The page could not be loaded or read, so nothing was measured |
When a target that passed in the previous run of the same runner gets worse, the weekly workflow opens an issue labeled stealth-regression. UNREACHABLE results never open issues.
Automated Results
Last updated 2026-10-26 from 7 runner(s). This section is generated by python -m benchmarks.stealth render from docs/stealth-results/, do not edit it by hand.
| Target | Category | linux-x64-headed-stable | linux-x64-headless-previous | linux-x64-headless-stable | macos-arm64-headed-stable | macos-arm64-headless-stable | windows-x64-headed-stable | windows-x64-headless-stable |
|---|---|---|---|---|---|---|---|---|
| Launch flags | Browser configuration | PASS | PASS | PASS | PASS | PASS | PASS | PASS |
| Local fingerprint probe | Automation tells | DEGRADED | DEGRADED | DEGRADED | PASS | PASS | PASS | PASS |
| Init script injection | Capability | PASS | PASS | PASS | PASS | PASS | PASS | PASS |
| Input fidelity | Interaction | PASS | PASS | PASS | PASS | PASS | PASS | PASS |
| Sannysoft | Bot detection suite | FAIL | FAIL | FAIL | PASS | PASS | PASS | PASS |
| Intoli | Headless detection | PASS | PASS | PASS | PASS | PASS | PASS | PASS |
| CreepJS | Deep fingerprinting | PASS | PASS | PASS | PASS | PASS | PASS | PASS |
| Cloudflare (nowsecure.nl) | JS challenge | PASS | PASS | PASS | PASS | PASS | PASS | PASS |
| TLS and HTTP/2 fingerprint | Network fingerprinting | PASS | PASS | PASS | PASS | PASS | PASS | PASS |
Runners
| Runner | Date | Chrome | nodriver | OS | Result |
|---|---|---|---|---|---|
| linux-x64-headed-stable | 2026-10-26 | 155.0.8059.39 | 0.47.0 | Linux-7.0.0-1017-azure-x86_64-with-glibc2.43 | 7 pass, 1 degraded, 1 fail |
| linux-x64-headless-previous | 2026-10-26 | 155.0.8059.84 | 0.47.0 | Linux-7.0.0-1017-azure-x86_64-with-glibc2.43 | 7 pass, 1 degraded, 1 fail |
| linux-x64-headless-stable | 2026-10-26 | 155.0.8059.39 | 0.47.0 | Linux-7.0.0-1017-azure-x86_64-with-glibc2.43 | 7 pass, 1 degraded, 1 fail |
| macos-arm64-headed-stable | 2026-10-26 | 155.0.8059.40 | 0.47.0 | macOS-26.6.2-arm64-arm-64bit | 9 pass |
| macos-arm64-headless-stable | 2026-10-26 | 156.0.8078.48 | 0.47.0 | macOS-26.6.2-arm64-arm-64bit | 9 pass |
| windows-x64-headed-stable | 2026-10-26 | 156.0.8078.48 | 0.47.0 | Windows-10-10.0.26100-SP0 | 9 pass |
| windows-x64-headless-stable | 2026-10-26 | 156.0.8078.48 | 0.47.0 | Windows-10-10.0.26100-SP0 | 9 pass |
Checks Not Passing
linux-x64-headed-stable
- Local fingerprint probe: WebGL renderer is hardware (minor)
- Sannysoft: WebGL Vendor, WebGL Renderer
linux-x64-headless-previous
- Local fingerprint probe: WebGL renderer is hardware (minor)
- Sannysoft: WebGL Renderer
linux-x64-headless-stable
- Local fingerprint probe: WebGL renderer is hardware (minor)
- Sannysoft: WebGL Renderer
History
Passing targets per run.
| Date | linux-x64-headed-stable | linux-x64-headless-previous | linux-x64-headless-stable | macos-arm64-headed-stable | macos-arm64-headless-stable | windows-x64-headed-stable | windows-x64-headless-stable |
|---|---|---|---|---|---|---|---|
| 2026-10-26 | 7/9 | 7/9 | 7/9 | 9/9 | 9/9 | 9/9 | 9/9 |
| 2026-10-19 | 7/9 | 7/9 | 7/9 | 9/9 | 9/9 | 9/9 | 9/9 |
| 2026-10-12 | 7/9 | 7/9 | 7/9 | 9/9 | 9/9 | 9/9 | 9/9 |
| 2026-10-10 | 7/9 | 7/9 | 7/9 | 9/9 | 9/9 | 9/9 | 9/9 |
Manual Snapshot (2026-02-10)
Browser: Chrome 144 (nodriver) Platform: Windows 10 (Win64) Mode: Non-headless
1. Cloudflare Challenge (nowsecure.nl)
Result: PASSED
The browser navigated through Cloudflare's bot challenge without any manual intervention. The page resolved to the "NOWSECURE BY NODRIVER" victory screen, confirming full bypass of Cloudflare's JavaScript challenge, TLS fingerprinting, and behavioral analysis.
2. CreepJS Fingerprint Analysis
Result: PASSED
CreepJS performs deep browser fingerprinting across dozens of vectors. Key findings:
| Detection Vector | Result |
|---|---|
| Headless detected | 0% |
| Stealth detected | 0% |
| Like headless | 25% (chromium flag, normal for real Chrome) |
| Privacy resistance | unknown (no fingerprint blocking detected) |
| Security resistance | unknown (clean) |
| Extension detected | unknown (none) |
Fingerprint consistency:
- GPU: NVIDIA GeForce RTX 2070 SUPER (real hardware)
- Platform: Windows 10 (64-bit), 16 cores, 8GB RAM
- Timezone: America/New_York (Eastern Standard Time)
- Language: en-US
- WebRTC: Real local/STUN candidates exposed
- Confidence: moderate (consistent fingerprint)
3. Sannysoft Bot Detection
Result: 20/20 TESTS PASSED
Every single detection vector returned "ok" or consistent values.
| Test | Result |
|---|---|
| PHANTOM_UA | ok |
| PHANTOM_PROPERTIES | ok |
| PHANTOM_ETSL | ok |
| PHANTOM_LANGUAGE | ok |
| PHANTOM_WEBSOCKET | ok |
| MQ_SCREEN | ok |
| PHANTOM_OVERFLOW | ok |
| PHANTOM_WINDOW_HEIGHT | ok |
| HEADCHR_UA | ok |
| HEADCHR_CHROME_OBJ | ok |
| HEADCHR_PERMISSIONS | ok |
| HEADCHR_PLUGINS | ok |
| HEADCHR_IFRAME | ok |
| CHR_DEBUG_TOOLS | ok |
| SELENIUM_DRIVER | ok |
| CHR_BATTERY | ok |
| CHR_MEMORY | ok |
| TRANSPARENT_PIXEL | ok |
| SEQUENTUM | ok |
| VIDEO_CODECS | ok |
Additional consistent properties:
- User Agent: Standard Chrome 144 UA string
- WebDriver: missing (passed - not detected as automated)
- Chrome object: present (passed - real Chrome API)
- Permissions: prompt (normal behavior)
- Plugins: 5 (standard Chrome PDF plugins)
- Languages: en-US, en
- WebGL Vendor: Google Inc. (NVIDIA)
- Broken Image: 16x16 (correct dimensions)
- Canvas fingerprints: All consistent hashes across contexts
4. Intoli Headless Detection (Round II)
Result: ALL TESTS PASSED
| Test | Result |
|---|---|
| User Agent (Old) | Chrome/144.0.0.0 (clean) |
| WebDriver (New) | missing (passed) |
| WebDriver Advanced | pass |
| Chrome (New) | present (passed) |
| Permissions (New) | prompt (normal) |
| Plugins Length (Old) | 5 (correct) |
| Plugins type | PluginArray (pass) |
| Languages (Old) | en-US, en (consistent) |
| WebGL Vendor | Google Inc. (NVIDIA) |
| WebGL Renderer | ANGLE (NVIDIA GeForce RTX 2070 SUPER) |
| Broken Image Dimensions | 16x16 (correct) |
5. X.com (Twitter) Login Wall
Result: PROFILE DATA READ BEHIND THE LOGIN WALL
Navigated to x.com/elonmusk, which shows a login modal over the content.
The profile data was already loaded behind the modal. After removing the
overlay through DOM manipulation, all profile data was readable:
- Display name, handle, bio, join date
- Follower/following counts (234.3M followers)
- Full tweet timeline with engagement metrics
- All data-testid attributes queryable
This shows logged-out reading of public profile data behind a login wall. No Arkose Labs or FunCaptcha challenge appeared in this flow, so this result says nothing about solving those challenges.
Summary
| Detection System | Type | Result |
|---|---|---|
| Cloudflare | JS Challenge + TLS | PASSED |
| CreepJS | Deep Fingerprinting | 0% stealth / 0% headless |
| Sannysoft | Bot Detection Suite | 20/20 tests passed |
| Intoli | Headless Detection | All tests passed |
| X.com | Login Wall | Profile data read behind the modal |
In this snapshot, stealth-browser-mcp running on nodriver was not flagged by any of the detection systems above. See the automated results for current numbers.