<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
<title>Browsa blog</title>
<subtitle>What we measure running cloud browsers for AI agents.</subtitle>
<link href="https://browsa.io/blog/feed.xml" rel="self" type="application/atom+xml"/>
<link href="https://browsa.io/blog" rel="alternate" type="text/html"/>
<id>https://browsa.io/blog</id>
<updated>2026-11-18T00:00:00Z</updated>
<author><name>Browsa team</name><uri>https://browsa.io/</uri></author>
<icon>https://browsa.io/browsa-icon-192.png</icon>
<entry>
<title>The post-quantum handshake that gives bots away</title>
<link href="https://browsa.io/blog/post-quantum-tls-handshake" rel="alternate" type="text/html"/>
<id>https://browsa.io/blog/post-quantum-tls-handshake</id>
<published>2026-11-18T00:00:00Z</published>
<updated>2026-11-18T00:00:00Z</updated>
<author><name>Browsa team</name></author>
<category term="Detection"/>
<summary>Chrome puts a post-quantum key share in its first TLS message. Clients that copy Chrome's User-Agent but not its handshake stand out before a page loads.</summary>
<content type="html">&lt;p&gt;&lt;strong&gt;TL;DR&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;
&lt;li&gt;Since 2024, Chrome's TLS ClientHello offers a hybrid post-quantum key share. Today it is X25519MLKEM768, codepoint 0x11EC, 1,216 bytes on its own.&lt;/li&gt;
&lt;li&gt;The server reads that message before any HTTP header or script. A client that says &quot;Chrome&quot; in its User-Agent but sends another library's handshake contradicts itself right there.&lt;/li&gt;
&lt;li&gt;Newer libraries now send the key share too, but the rest of their handshake still differs. JA4 sums that up in one string.&lt;/li&gt;
&lt;li&gt;The fix: use a real browser engine, or a TLS library that copies the current Chrome, and check the result.&lt;/li&gt;
&lt;/ul&gt;&lt;p&gt;A bot check does not need JavaScript to spot most scripted clients. The first message of every HTTPS connection, the TLS ClientHello, is sent before the request line and says a lot about the software that built it. Since 2024, Chrome's version of that message carries something many HTTP libraries did not have for a long time: a post-quantum key share.&lt;/p&gt;

&lt;h2 id=&quot;what-chrome-sends&quot;&gt;What Chrome sends now&lt;/h2&gt;
&lt;p&gt;TLS 1.3 lets a client guess the key exchange the server will pick and send its half of the key in the first message, so the handshake needs one round trip instead of two. Chrome now guesses a hybrid of classic X25519 and a post-quantum scheme, and it got there in three steps:&lt;/p&gt;
&lt;div class=&quot;ds-table-wrap&quot;&gt;&lt;table class=&quot;ds-table&quot;&gt;
&lt;thead&gt;&lt;tr&gt;&lt;th scope=&quot;col&quot;&gt;When&lt;/th&gt;&lt;th scope=&quot;col&quot;&gt;What changed&lt;/th&gt;&lt;th scope=&quot;col&quot;&gt;Source&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;&lt;td&gt;Aug 2023&lt;/td&gt;&lt;td&gt;Chrome announces hybrid Kyber key agreement, X25519Kyber768, starting as an option.&lt;/td&gt;&lt;td&gt;&lt;a href=&quot;https://blog.chromium.org/2023/08/protecting-chrome-traffic-with-hybrid.html&quot;&gt;Chromium blog&lt;/a&gt;&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Apr 2024&lt;/td&gt;&lt;td&gt;Version 124 turns it on by default on desktop (codepoint 0x6399).&lt;/td&gt;&lt;td&gt;&lt;a href=&quot;https://groups.google.com/a/chromium.org/g/blink-dev/c/6xfaov3Z4yo&quot;&gt;Intent to ship&lt;/a&gt;&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Nov 2024&lt;/td&gt;&lt;td&gt;Version 131 drops Kyber and offers the standard ML-KEM hybrid, X25519MLKEM768, codepoint 0x11EC.&lt;/td&gt;&lt;td&gt;&lt;a href=&quot;https://security.googleblog.com/2024/09/a-new-path-for-kyber-on-web.html&quot;&gt;Google Security blog&lt;/a&gt;&lt;/td&gt;&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;
&lt;p&gt;The IETF draft that defines the hybrid puts the client share at &lt;a href=&quot;https://datatracker.ietf.org/doc/draft-ietf-tls-ecdhe-mlkem/&quot;&gt;1,216 bytes&lt;/a&gt;: 1,184 for ML-KEM-768 and 32 for X25519. A classic X25519 share is 32 bytes. That one field pushes the ClientHello past what fits in a single packet, which is why some servers broke when Chrome turned it on: they expected the whole hello in one &lt;code&gt;read()&lt;/code&gt;. The &lt;a href=&quot;https://tldr.fail/&quot;&gt;tldr.fail&lt;/a&gt; page documents that bug.&lt;/p&gt;

&lt;h2 id=&quot;same-ua&quot;&gt;Same User-Agent, different handshake&lt;/h2&gt;
&lt;p&gt;A scraper that sends a Chrome User-Agent from Python, Go or curl changes one HTTP header. The TLS library underneath still writes its own ClientHello: its own cipher list, its own extensions, its own key share. The server sees the handshake first and the header second.&lt;/p&gt;
&lt;p&gt;We pointed six clients at &lt;a href=&quot;https://tls.peet.ws/api/all&quot;&gt;tls.peet.ws&lt;/a&gt;, which echoes back the ClientHello it received, on 17 November 2026:&lt;/p&gt;
&lt;div class=&quot;ds-table-wrap&quot;&gt;&lt;table class=&quot;ds-table bl-mono-table&quot;&gt;
&lt;thead&gt;&lt;tr&gt;&lt;th scope=&quot;col&quot;&gt;Client&lt;/th&gt;&lt;th scope=&quot;col&quot;&gt;Key share&lt;/th&gt;&lt;th scope=&quot;col&quot;&gt;JA4&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;&lt;td&gt;Google Chrome, current stable, macOS&lt;/td&gt;&lt;td&gt;X25519MLKEM768, X25519&lt;/td&gt;&lt;td&gt;&lt;code&gt;t13d1517h2_8daaf6152771_cb7bf5808d99&lt;/code&gt;&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Phantom, Browsa production&lt;/td&gt;&lt;td&gt;X25519MLKEM768, X25519&lt;/td&gt;&lt;td&gt;&lt;code&gt;t13d1517h2_8daaf6152771_cb7bf5808d99&lt;/code&gt;&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;curl 8.7.1, macOS system build (LibreSSL)&lt;/td&gt;&lt;td&gt;X25519 only&lt;/td&gt;&lt;td&gt;&lt;code&gt;t13d497h2_0d8feac7bc37_7395dae3b2f3&lt;/code&gt;&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Python requests 2.32, OpenSSL 3.6&lt;/td&gt;&lt;td&gt;X25519MLKEM768, X25519&lt;/td&gt;&lt;td&gt;&lt;code&gt;t13d1712h1_ab0a1bf427ad_882d495ac381&lt;/code&gt;&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Go 1.26 &lt;code&gt;net/http&lt;/code&gt;&lt;/td&gt;&lt;td&gt;X25519MLKEM768, X25519&lt;/td&gt;&lt;td&gt;&lt;code&gt;t13d1312h2_f57a46bbacb6_ab7e3b40a677&lt;/code&gt;&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Node 26 &lt;code&gt;fetch&lt;/code&gt;&lt;/td&gt;&lt;td&gt;X25519MLKEM768, X25519&lt;/td&gt;&lt;td&gt;&lt;code&gt;t13d5212h1_b262b3658495_8e6e362c5eac&lt;/code&gt;&lt;/td&gt;&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;
&lt;p&gt;Two things stand out. First, the missing key share is now a sign of an &lt;em&gt;old&lt;/em&gt; stack, not of every library. OpenSSL added X25519MLKEM768 to its default key shares in &lt;a href=&quot;https://openssl-library.org/news/openssl-3.5-notes/&quot;&gt;3.5.0&lt;/a&gt; (April 2025). Go offered the Kyber draft by default in &lt;a href=&quot;https://go.dev/doc/go1.23&quot;&gt;1.23&lt;/a&gt; and switched to ML-KEM in &lt;a href=&quot;https://go.dev/doc/go1.24&quot;&gt;1.24&lt;/a&gt;. Python linked against OpenSSL 3.4 or older, Go before 1.23, and curl builds like the macOS one above still send a classic share only. Go 1.23 sends the 0x6399 draft that Chrome stopped offering in 2024, which is a mismatch of its own.&lt;/p&gt;
&lt;p&gt;Second, a matching key share is not a matching handshake. Python, Go and Node all offer ML-KEM now, and none of them looks like Chrome: different cipher counts, different extensions. Python requests offers only HTTP/1.1, and Node lists HTTP/1.1 before HTTP/2. Chrome and Phantom produce the same JA4.&lt;/p&gt;

&lt;h2 id=&quot;ja4&quot;&gt;JA4 in one minute&lt;/h2&gt;
&lt;p&gt;&lt;a href=&quot;https://github.com/FoxIO-LLC/ja4/blob/main/technical_details/JA4.md&quot;&gt;JA4&lt;/a&gt;, from FoxIO, turns a ClientHello into a short string with three parts. Take Chrome's:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;code&gt;t13d1517h2&lt;/code&gt;: TLS over TCP, version 1.3, a domain name in SNI, 15 cipher suites, 17 extensions, and &lt;code&gt;h2&lt;/code&gt; as the first ALPN value.&lt;/li&gt;
&lt;li&gt;&lt;code&gt;8daaf6152771&lt;/code&gt;: a truncated SHA-256 of the cipher suites, sorted.&lt;/li&gt;
&lt;li&gt;&lt;code&gt;cb7bf5808d99&lt;/code&gt;: a truncated SHA-256 of the extensions, sorted, plus the signature algorithms in their sent order.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;JA4 ignores GREASE values and sorts the lists, so a client that shuffles its extension order still gets one stable value. It does not hash the key share bytes, but the extension list and counts move whenever a library's defaults move, and a Chrome User-Agent next to a JA4 that no Chrome release sends is a contradiction a server can check before it serves a byte.&lt;/p&gt;
&lt;p&gt;The handshake also changes between Chrome releases. A TLS profile copied from mid-2024 Chrome offers the 0x6399 Kyber share, which Chrome stopped sending in version 131. A frozen copy of Chrome becomes a fingerprint of its own.&lt;/p&gt;

&lt;h2 id=&quot;check&quot;&gt;Check your own client&lt;/h2&gt;
&lt;p&gt;Run these from the same machine, library and proxy your agent uses. Replace &lt;code&gt;curl&lt;/code&gt; with your client.&lt;/p&gt;
&lt;div class=&quot;ds-code&quot;&gt;&lt;pre&gt;&lt;code&gt;# JA4 of the ClientHello your client sent
curl -s https://tls.peet.ws/api/all | python3 -c &quot;import sys, json; print(json.load(sys.stdin)['tls']['ja4'])&quot;

# The key exchange Cloudflare negotiated with you
curl -s https://www.cloudflare.com/cdn-cgi/trace | grep kex&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;
&lt;p&gt;Chrome and Phantom print &lt;code&gt;kex=X25519MLKEM768&lt;/code&gt; on the second line. The macOS system curl prints &lt;code&gt;kex=X25519&lt;/code&gt;.&lt;/p&gt;
&lt;p&gt;To check a whole browser rather than one HTTP client, open &lt;a href=&quot;https://browsa.io/detect&quot;&gt;browsa.io/detect&lt;/a&gt; in your agent's browser. It checks TLS, HTTP/2 and the IP address along with what page scripts can see.&lt;/p&gt;

&lt;h2 id=&quot;what-to-do&quot;&gt;What to do about it&lt;/h2&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;Use a real browser engine.&lt;/strong&gt; A browser sends its own handshake, so there is nothing to keep in sync. Phantom, Browsa's own browser, uses the same TLS code as Chrome: in the capture above its JA4 matched Chrome's exactly, and so did its HTTP/2 settings fingerprint.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Or use a library that copies the current Chrome.&lt;/strong&gt; Projects such as curl-impersonate, curl_cffi and uTLS rebuild Chrome's ClientHello outside the browser. Pick a profile for the Chrome version your User-Agent claims, update both together, and run the check above after each update.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Keep the story consistent.&lt;/strong&gt; The handshake, the HTTP/2 settings, the headers and the User-Agent all describe one client. Changing only the header is the mismatch these checks look for.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;None of this gets a client past every bot check. The IP address, the browser's JavaScript surface and its behavior all count too. But the handshake is the first thing a server sees, and it is the cheapest one to get right.&lt;/p&gt;&lt;h2&gt;Sources&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://blog.chromium.org/2023/08/protecting-chrome-traffic-with-hybrid.html&quot;&gt;Protecting Chrome Traffic with Hybrid Kyber KEM&lt;/a&gt;, Chromium blog, August 2023&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://groups.google.com/a/chromium.org/g/blink-dev/c/6xfaov3Z4yo&quot;&gt;Intent to Ship: X25519Kyber768 key encapsulation for TLS on Desktop&lt;/a&gt;, blink-dev, March 2024&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://security.googleblog.com/2024/09/a-new-path-for-kyber-on-web.html&quot;&gt;A new path for Kyber on the web&lt;/a&gt;, Google Online Security blog, September 2024&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://datatracker.ietf.org/doc/draft-ietf-tls-ecdhe-mlkem/&quot;&gt;draft-ietf-tls-ecdhe-mlkem&lt;/a&gt;, IETF: codepoint 4588 (0x11EC), client share 1,216 bytes&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://tldr.fail/&quot;&gt;tldr.fail&lt;/a&gt;: servers that read a large ClientHello in one call&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://go.dev/doc/go1.23&quot;&gt;Go 1.23&lt;/a&gt; and &lt;a href=&quot;https://go.dev/doc/go1.24&quot;&gt;Go 1.24&lt;/a&gt; release notes, crypto/tls&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://openssl-library.org/news/openssl-3.5-notes/&quot;&gt;OpenSSL 3.5 release notes&lt;/a&gt;, 8 April 2025&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://github.com/FoxIO-LLC/ja4/blob/main/technical_details/JA4.md&quot;&gt;JA4 technical details&lt;/a&gt;, FoxIO&lt;/li&gt;
&lt;li&gt;Our captures: &lt;a href=&quot;https://tls.peet.ws/api/all&quot;&gt;tls.peet.ws/api/all&lt;/a&gt; and Cloudflare's &lt;code&gt;/cdn-cgi/trace&lt;/code&gt;, 17 November 2026&lt;/li&gt;&lt;/ol&gt;</content>
</entry>
<entry>
<title>Captchas, measured: only what each site's server accepted</title>
<link href="https://browsa.io/blog/captchas-measured" rel="alternate" type="text/html"/>
<id>https://browsa.io/blog/captchas-measured</id>
<published>2026-11-18T00:00:00Z</published>
<updated>2026-11-18T00:00:00Z</updated>
<author><name>Browsa team</name></author>
<category term="Captchas"/>
<summary>216 of 223 captcha attempts accepted by the site's own server, on production browsers, misses counted. How we measure, and what we don't cover yet.</summary>
<content type="html">&lt;p&gt;&lt;strong&gt;TL;DR&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;
&lt;li&gt;13 captcha types are solved inside Phantom today: 216 of 223 attempts were accepted by the site's own server.&lt;/li&gt;
&lt;li&gt;Every attempt runs on a production browser, against a challenge that changes on each load. Misses stay in the total, and the set is re-run every month.&lt;/li&gt;
&lt;li&gt;Limits: AWS WAF and hCaptcha were measured through an ISP proxy only. reCAPTCHA Enterprise, DataDome and Arkose are not covered yet.&lt;/li&gt;
&lt;/ul&gt;&lt;p&gt;Most captcha claims are one percentage with no method behind it. Ours is a count you can check: 216 of 223 attempts across 13 captcha types, each accepted or refused by the server of the site that showed the captcha. This post explains how those attempts are run and counted, and where the numbers stop. The live table is on &lt;a href=&quot;https://browsa.io/captcha&quot;&gt;/captcha&lt;/a&gt;.&lt;/p&gt;

&lt;h2 id=&quot;numbers&quot;&gt;The numbers&lt;/h2&gt;
&lt;p&gt;These are the rows switched on for customers, as published on 17 November 2026. Each row is its own set of attempts on production browsers.&lt;/p&gt;
&lt;div class=&quot;ds-table-wrap&quot;&gt;&lt;table class=&quot;ds-table bl-score&quot;&gt;
&lt;thead&gt;&lt;tr&gt;&lt;th scope=&quot;col&quot;&gt;Captcha&lt;/th&gt;&lt;th scope=&quot;col&quot;&gt;Accepted&lt;/th&gt;&lt;th scope=&quot;col&quot;&gt;Median&lt;/th&gt;&lt;th scope=&quot;col&quot;&gt;Exit&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;&lt;td&gt;reCAPTCHA v2, 3 x 3 image grid&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:100.0%&quot;&gt;&lt;/span&gt;16 of 16&lt;/td&gt;&lt;td&gt;13.6 s&lt;/td&gt;&lt;td&gt;Data center&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;reCAPTCHA v2, 4 x 4 single image&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:94.4%&quot;&gt;&lt;/span&gt;17 of 18&lt;/td&gt;&lt;td&gt;16.1 s&lt;/td&gt;&lt;td&gt;Data center&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;reCAPTCHA v2, grid with reloading tiles&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:93.3%&quot;&gt;&lt;/span&gt;14 of 15&lt;/td&gt;&lt;td&gt;70.6 s&lt;/td&gt;&lt;td&gt;Data center and US ISP proxy&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;Image text, type the characters&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:93.3%&quot;&gt;&lt;/span&gt;28 of 30&lt;/td&gt;&lt;td&gt;6.2 s&lt;/td&gt;&lt;td&gt;Data center&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;GeeTest v3, slide puzzle&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:93.8%&quot;&gt;&lt;/span&gt;15 of 16&lt;/td&gt;&lt;td&gt;12.9 s&lt;/td&gt;&lt;td&gt;Data center and US ISP proxy&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;GeeTest v4, slide puzzle&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:93.8%&quot;&gt;&lt;/span&gt;15 of 16&lt;/td&gt;&lt;td&gt;12.7 s&lt;/td&gt;&lt;td&gt;Data center and US ISP proxy&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;AWS WAF, captcha grid&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:100.0%&quot;&gt;&lt;/span&gt;16 of 16&lt;/td&gt;&lt;td&gt;15.1 s&lt;/td&gt;&lt;td&gt;US ISP proxy&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;hCaptcha, picture and drag puzzles&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:100.0%&quot;&gt;&lt;/span&gt;16 of 16&lt;/td&gt;&lt;td&gt;17.5 s&lt;/td&gt;&lt;td&gt;US ISP proxy&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;GeeTest v4, IconCrush (swap to make three)&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:100.0%&quot;&gt;&lt;/span&gt;16 of 16&lt;/td&gt;&lt;td&gt;7.9 s&lt;/td&gt;&lt;td&gt;Data center&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;GeeTest v4, gobang (move one to make five)&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:100.0%&quot;&gt;&lt;/span&gt;16 of 16&lt;/td&gt;&lt;td&gt;8.1 s&lt;/td&gt;&lt;td&gt;Data center&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;GeeTest v4, click the icons in order&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:93.8%&quot;&gt;&lt;/span&gt;15 of 16&lt;/td&gt;&lt;td&gt;14.1 s&lt;/td&gt;&lt;td&gt;Data center&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;GeeTest v4, pick the shape that matches&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:100.0%&quot;&gt;&lt;/span&gt;16 of 16&lt;/td&gt;&lt;td&gt;18.9 s&lt;/td&gt;&lt;td&gt;Data center&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;GeeTest v4, pick the photos that match&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;bl-bar&quot; style=&quot;--p:100.0%&quot;&gt;&lt;/span&gt;16 of 16&lt;/td&gt;&lt;td&gt;10.5 s&lt;/td&gt;&lt;td&gt;Data center&lt;/td&gt;&lt;/tr&gt;
&lt;/tbody&gt;
&lt;tfoot&gt;&lt;tr&gt;&lt;th scope=&quot;row&quot;&gt;Total&lt;/th&gt;&lt;td&gt;216 of 223&lt;/td&gt;&lt;td&gt;&lt;/td&gt;&lt;td&gt;&lt;/td&gt;&lt;/tr&gt;&lt;/tfoot&gt;
&lt;/table&gt;&lt;/div&gt;
&lt;p&gt;Three more checks pass with no solver at all, because the browser itself is accepted: reCAPTCHA v3 (score 0.9 on 5 of 5), Cloudflare's full-page check (5 of 5, one click) and Cloudflare Turnstile (15 of 15 tokens accepted by Cloudflare's &lt;code&gt;siteverify&lt;/code&gt; on a widget we run for this purpose). They are not in the 216.&lt;/p&gt;
&lt;p&gt;The reloading-tiles row is slow because of how that puzzle works: Google takes about 5 seconds to replace each clicked tile and asks for several rounds. Its one miss ran out of rounds at 134 s.&lt;/p&gt;

&lt;h2 id=&quot;method&quot;&gt;How we count&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Production browsers.&lt;/strong&gt; Attempts run on the same Phantom browsers and the same solver that customers use, not a lab copy.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;A new challenge every load.&lt;/strong&gt; Before a number is published, we check that the page serves a different challenge on each load and uses a real site key. A test key always passes, and a fixed picture can be memorized.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The site's server decides.&lt;/strong&gt; A pass means the server behind the captcha accepted the answer: Google's verification on its demo, Cloudflare's &lt;code&gt;siteverify&lt;/code&gt;, BotDetect's own check, GeeTest's server through its demo's login check, or AWS setting its WAF token and the page behind it loading. A token that the solver produced but no server accepted is not a pass.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Misses stay in.&lt;/strong&gt; Each attempt uses a fresh browser. Every attempt that reached the captcha type counts, including the slow ones. When a solver is changed, the row shows the run of the version that ships, and earlier runs stay in its log.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Re-run monthly.&lt;/strong&gt; The same benchmark runs again each month and the dated rows change with it. The script, &lt;a href=&quot;https://browsa.io/media/captcha/capbench.js&quot;&gt;capbench.js&lt;/a&gt;, and each row's attempt log can be downloaded from &lt;a href=&quot;https://browsa.io/captcha&quot;&gt;/captcha&lt;/a&gt;.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;A type is switched on for customers only after at least 15 production attempts with at least 90% accepted. Press and hold (HUMAN) is at 11 of 12 and is marked Beta until it has more attempts.&lt;/p&gt;
&lt;p&gt;We got this wrong once. An early Turnstile row of 5 of 5 came from a demo that runs Cloudflare's test key, which always passes. We replaced it with the server-verified 15 of 15 and wrote the correction on &lt;a href=&quot;https://browsa.io/captcha&quot;&gt;/captcha&lt;/a&gt;. That mistake is why the &quot;real site key&quot; check is in the list above.&lt;/p&gt;

&lt;h2 id=&quot;limits&quot;&gt;Limits&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Exit address.&lt;/strong&gt; AWS WAF and hCaptcha were measured through a US ISP proxy only. From our data-center exit, the AWS test page answered 403 before any puzzle on 8 of 8 visits, and hcaptcha.com blocks the visit too. No solver can fix a visit that never reaches the puzzle.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Sample size.&lt;/strong&gt; 15 to 30 attempts per type is small. 16 of 16 does not mean 100%: by the rule of three, the true miss rate is likely below about 19% (3 divided by 16), not zero.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Demo pages.&lt;/strong&gt; Most rows use the vendor's own demo or a public demo page. A site can set its captcha harder than the demo.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Not covered yet.&lt;/strong&gt; reCAPTCHA Enterprise passed 7 of 22 in tests and is not offered. DataDome was 0 of 5 on a protected site. Arkose is not measured: there is no public demo with a server verdict.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2 id=&quot;claims&quot;&gt;How claims usually look&lt;/h2&gt;
&lt;p&gt;Search for captcha solving and you will find &quot;99% success&quot; or &quot;solves in seconds&quot; on many landing pages. Few of them say which captcha, on which site, how many attempts, who decided what counted as a pass, from which IP address, or when. Without those, a percentage cannot be checked or compared.&lt;/p&gt;
&lt;p&gt;If you are comparing providers, including us, ask for:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;The exact captcha type and variant, such as reCAPTCHA v2 4 x 4, not just &quot;reCAPTCHA&quot;.&lt;/li&gt;
&lt;li&gt;Whether the site key was real and the challenge changed on each load.&lt;/li&gt;
&lt;li&gt;Whose server accepted each answer.&lt;/li&gt;
&lt;li&gt;Attempts and misses, not only a rate.&lt;/li&gt;
&lt;li&gt;The exit address type and the date of the run.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;That is the format of every row on &lt;a href=&quot;https://browsa.io/captcha&quot;&gt;/captcha&lt;/a&gt;. When a number goes down at the next monthly run, the page will show the lower number.&lt;/p&gt;&lt;h2&gt;Sources&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://browsa.io/captcha&quot;&gt;Browsa captcha results&lt;/a&gt;: every row, its demo recording and its attempt log&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://browsa.io/media/captcha/capbench.js&quot;&gt;capbench.js&lt;/a&gt;: the benchmark script&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://developers.cloudflare.com/turnstile/get-started/server-side-validation/&quot;&gt;Turnstile server-side validation&lt;/a&gt;, Cloudflare docs: how &lt;code&gt;siteverify&lt;/code&gt; accepts a token&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://developers.cloudflare.com/turnstile/troubleshooting/testing/&quot;&gt;Turnstile testing keys&lt;/a&gt;, Cloudflare docs: test keys that always pass&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://en.wikipedia.org/wiki/Rule_of_three_(statistics)&quot;&gt;Rule of three&lt;/a&gt;, a 95% bound when no misses are seen&lt;/li&gt;&lt;/ol&gt;</content>
</entry>
<entry>
<title>One DOM query fails Cloudflare's checkbox</title>
<link href="https://browsa.io/blog/dom-query-fails-cloudflare-check" rel="alternate" type="text/html"/>
<id>https://browsa.io/blog/dom-query-fails-cloudflare-check</id>
<published>2026-11-18T00:00:00Z</published>
<updated>2026-11-18T00:00:00Z</updated>
<author><name>Browsa team</name></author>
<category term="Engineering"/>
<summary>A probe that queried the DOM every 0.5 s made Cloudflare's checkbox never clear. Reading the tab title over CDP fixed it. What we measured, and what to do.</summary>
<content type="html">&lt;p&gt;&lt;strong&gt;TL;DR&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;
&lt;li&gt;From a data-center address, Cloudflare shows its checkbox page. In Phantom, one click on the checkbox passes it.&lt;/li&gt;
&lt;li&gt;While a probe queried the page's DOM every 0.5 s, the same click never cleared. With a probe that only read &lt;code&gt;typeof window._cf_chl_opt&lt;/code&gt;, it cleared in 1 s.&lt;/li&gt;
&lt;li&gt;Don't run code that touches the page while a challenge may be up. Read the tab title from the browser over CDP (&lt;code&gt;Target.getTargets&lt;/code&gt;) instead.&lt;/li&gt;
&lt;/ul&gt;&lt;p&gt;An agent that checks whether a page is ready can be the reason the page never becomes ready. On Cloudflare's full-page check, our own status probe was enough to make a valid click fail, every time. This post covers what we measured on production on 12 November 2026 and how the runner works now. It is not a way around Cloudflare: the check is built to let a real browser through after one click, and the lesson is how not to break that.&lt;/p&gt;

&lt;h2 id=&quot;setup&quot;&gt;The setup&lt;/h2&gt;
&lt;p&gt;Browsa runs agent tasks in Phantom, Browsa's own browser. Without a proxy, a browser exits from our data-center address, and on many Cloudflare sites that address alone brings up the &quot;Verify you are human&quot; page. The runner's job on that page is short: notice the check, click the checkbox once, and notice when the real page has loaded.&lt;/p&gt;
&lt;p&gt;With the Phantom build shipped that day, one click on the checkbox passed. With the previous build, the same click from the same address was refused. The address decides whether the check appears; the browser decides whether the click is accepted.&lt;/p&gt;

&lt;h2 id=&quot;probe&quot;&gt;What broke it&lt;/h2&gt;
&lt;p&gt;To notice the check and its end, the runner asked the page about its state every 0.5 s through CDP's &lt;code&gt;Runtime.evaluate&lt;/code&gt;. We tried two versions of that question on the same page, same browser, same click:&lt;/p&gt;
&lt;div class=&quot;ds-table-wrap&quot;&gt;&lt;table class=&quot;ds-table&quot;&gt;
&lt;thead&gt;&lt;tr&gt;&lt;th scope=&quot;col&quot;&gt;Probe, every 0.5 s&lt;/th&gt;&lt;th scope=&quot;col&quot;&gt;After one click&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;&lt;td&gt;&lt;code&gt;document.querySelector(...)&lt;/code&gt; and &lt;code&gt;document.title&lt;/code&gt;&lt;/td&gt;&lt;td&gt;Never cleared&lt;/td&gt;&lt;/tr&gt;
&lt;tr&gt;&lt;td&gt;&lt;code&gt;typeof window._cf_chl_opt&lt;/code&gt; only&lt;/td&gt;&lt;td&gt;Cleared in 1 s&lt;/td&gt;&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;
&lt;p&gt;The first probe only read. It changed nothing on the page, yet the check never accepted the click while it ran. We do not know which signal Cloudflare reads here, and Cloudflare does not document it. We only know the outcome, and that it repeated.&lt;/p&gt;

&lt;h2 id=&quot;leftover&quot;&gt;The variable that outlives the check&lt;/h2&gt;
&lt;p&gt;The light probe has a catch. &lt;code&gt;window._cf_chl_opt&lt;/code&gt; is defined on the check page, but it stays defined after the check passes, because the real page is swapped in without a full navigation. A probe that trusts it alone will report a check that is already gone.&lt;/p&gt;
&lt;p&gt;The fix is to ask the browser, not the page. CDP's &lt;code&gt;Target.getTargets&lt;/code&gt; returns every tab's URL and title from the browser process, without running anything inside the page. The check page has a recognizable title, so the runner treats a yes from the variable as true only while the tab title still matches:&lt;/p&gt;
&lt;div class=&quot;ds-code&quot;&gt;&lt;pre&gt;&lt;code&gt;import re

# Cloudflare translates the title; add the languages your browsers use.
CHECK_TITLE = re.compile(r&quot;^\s*(just a moment|attention required|nur einen moment)&quot;, re.I)

async def check_page_is_up(browser, url):
    &quot;&quot;&quot;Read tab titles from the browser. Nothing runs inside the page.&quot;&quot;&quot;
    cdp = await browser.new_browser_cdp_session()  # Playwright, async API
    infos = (await cdp.send(&quot;Target.getTargets&quot;))[&quot;targetInfos&quot;]
    base = url.split(&quot;#&quot;)[0]
    titles = [t[&quot;title&quot;] for t in infos
              if t[&quot;type&quot;] == &quot;page&quot; and t[&quot;url&quot;].split(&quot;#&quot;)[0] == base and t[&quot;title&quot;].strip()]
    if not titles:
        return None  # still loading: no answer yet
    return any(CHECK_TITLE.match(t) for t in titles)&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;
&lt;p&gt;An empty title mid-load proves nothing, so it returns no answer rather than &quot;passed&quot;. We have seen the German title, &quot;Nur einen Moment...&quot;, on browsers set to a German locale.&lt;/p&gt;

&lt;h2 id=&quot;timing&quot;&gt;When to click&lt;/h2&gt;
&lt;p&gt;Timing mattered too. On a busy browser, a click sent about 7 s after the check page appeared did nothing: the widget was still starting. Later clicks worked, and every accepted click cleared the page within 2 s.&lt;/p&gt;
&lt;figure class=&quot;bl-timeline&quot; aria-label=&quot;Timeline of one Cloudflare check&quot;&gt;
&lt;ol&gt;
&lt;li&gt;&lt;b&gt;0 s&lt;/b&gt;&lt;span&gt;Check page appears&lt;/span&gt;&lt;/li&gt;
&lt;li class=&quot;is-miss&quot;&gt;&lt;b&gt;about 7 s&lt;/b&gt;&lt;span&gt;A click here can be ignored&lt;/span&gt;&lt;/li&gt;
&lt;li class=&quot;is-go&quot;&gt;&lt;b&gt;9 s&lt;/b&gt;&lt;span&gt;The runner clicks once&lt;/span&gt;&lt;/li&gt;
&lt;li class=&quot;is-go&quot;&gt;&lt;b&gt;within 2 s&lt;/b&gt;&lt;span&gt;Real page loads&lt;/span&gt;&lt;/li&gt;
&lt;/ol&gt;
&lt;figcaption&gt;Measured on production, 12 November 2026. If a click is ignored, the runner waits 12 s and tries again, up to 3 clicks.&lt;/figcaption&gt;
&lt;/figure&gt;
&lt;p&gt;So the runner now waits 9 s before the first click, gives each click 12 s to clear, and stops after 3 clicks. On &lt;a href=&quot;https://browsa.io/captcha&quot;&gt;/captcha&lt;/a&gt;, Cloudflare's full-page check passed 5 of 5 times from the data-center address, with a median of 13.5 s.&lt;/p&gt;

&lt;h2 id=&quot;lessons&quot;&gt;Lessons for agent builders&lt;/h2&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;Keep your code out of the challenge page.&lt;/strong&gt; Anything that queries the DOM on a timer is suspect while a check may be up: readiness checks, element indexers, page scripts that build an accessibility snapshot. We measured only our own probe, so test yours.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Read state from the browser.&lt;/strong&gt; Tab titles and URLs from &lt;code&gt;Target.getTargets&lt;/code&gt; cost nothing on the page side.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Don't trust one signal.&lt;/strong&gt; A page variable can outlive the page it describes. Confirm it with a second source.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Click once, then wait.&lt;/strong&gt; Give the widget time to start, and give a click time to work before you send another.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;Browsa's hosted tasks do all of this already. If you drive your own Playwright or Puppeteer code on a Browsa browser, the same rules apply to your code.&lt;/p&gt;&lt;h2&gt;Sources&lt;/h2&gt;&lt;ol&gt;&lt;li&gt;&lt;a href=&quot;https://chromedevtools.github.io/devtools-protocol/tot/Target/#method-getTargets&quot;&gt;Target.getTargets&lt;/a&gt;, Chrome DevTools Protocol&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://playwright.dev/python/docs/api/class-browser#browser-new-browser-cdp-session&quot;&gt;Browser.new_browser_cdp_session&lt;/a&gt;, Playwright for Python&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://developers.cloudflare.com/cloudflare-challenges/&quot;&gt;Cloudflare challenges&lt;/a&gt;, Cloudflare docs&lt;/li&gt;
&lt;li&gt;&lt;a href=&quot;https://browsa.io/captcha&quot;&gt;Browsa captcha results&lt;/a&gt;: the full-page check row&lt;/li&gt;&lt;/ol&gt;</content>
</entry>
</feed>
