We offered $100 to anyone who could find a site that blocks our browser agent. We could test 82 of the sites people sent. Seven blocked us on every run, and all seven had the same problem: a captcha our solver didn't support at the time.
Six more sites blocked us only because we came from the US, and opened from a proxy in their own country.
Our fingerprint is good (bot detection, stealth benchmark), and it still hits walls.
Look at the wall first
Most blocks are one of these:
- A puzzle. Solve it and you're in.
- A banned IP. Solving does nothing. You need a new exit.
- A sensor flag. A script on the page didn't like how you behaved. A new IP does nothing. Behave differently.
- A country wall. The site serves one country and refuses the rest.
Our agents work inside people's accounts, so every fix below has to keep the login alive.
Which vendors can a solver beat?
From Nov 8 to 15, our US cloud browsers met 1,394,724 captchas. 79.2% were solved automatically. (An encounter is one session meeting one vendor's challenge, staff excluded. Sessions with the solver off send no events, so the real count is higher.)
DataDome and Akamai look bad, and a solver can't fix either. Most DataDome misses are pages closed mid-solve. Take those out and it's 82.3%. Akamai is mostly a 403 that our driver reloads once, and if it survives the reload, the IP is banned or a sensor flagged us.
A solver that never quits is a bug
In June, our PerimeterX solver succeeded on 1.7% of 110,208 detections, because 57% of those "captchas" were JavaScript files containing the vendor's name. DataDome had a similar problem: roughly 40% of its detections were script files too.
The worse case is an IP block that shows a captcha. Google's "unusual traffic" page shows a reCAPTCHA, and solving it doesn't lift the block. In 16 hours of July traffic, 73% of sessions that got a valid token were challenged again. The solver now spots that page, stops, and tells the agent to rotate the IP. DataDome gets the same treatment: reload once, then fail with "rotate IP, not retryable". Before that, one site cost us 106 solver starts and 106 failures in an hour.
The agent has to wait
In July our older agent capped its wait tool at 3 seconds and called a URL "captcha-blocked" after one wait and one reload. On 43 sign-up forms it ended captcha-blocked on 17. The newer agent, which waits, ended blocked on 10. We raised the cap to 30 seconds and haven't re-measured.
Default agents keep the solver on and just need to stop clicking:
When a captcha appears, stop driving and wait: a solve normally takes 20-60s. Re-check the page every ~5s for up to 2 minutes
Our fastest agent runs with the solver off, because the always-on detector kept the main thread 32-70% busy on a heavy flight-search page (the whole story). So it gets a solve_captcha tool that turns the solver on, waits up to 20 s for a captcha to show up and up to 95 s for it to finish, then turns it off again.
On staging it called the tool unprompted and solved hCaptcha 6 of 6 times, 21-30 s each. The same advice as a paragraph of prose went 0 of 3. Three tries is nothing, but I'd still give the agent a tool before a paragraph.
On Stealth Bench V2 our anti-detect browser got through 90.5% of 200 visits to 100 hard sites with the solver on and 84.0% with it off, so the solver is worth 6.5 points. Chrome with a US proxy got 14%. We picked those sites because our browser had got through and stock Chrome with a US proxy hadn't, so it flatters us. More in captcha solving for AI agents.
Change the IP without losing the login
A fresh browser fixes a burned IP and throws away every tab and cookie. So we gave agents a tool that swaps the exit under a running browser:
refresh_connection({ country: "jp", reason: "Site says it is not available in your region" })The tool shipped on Nov 8, and agents called it about 10 times per 100 runs of our current agent (some runs may not have had it yet). It doesn't work on customers' own proxies, since we can't rotate an IP we don't own, so it's hidden there.
It came out of a mistake. Our first "refresh" was a browser swap, and when the agent guessed wrong about why it was blocked, the swap destroyed state it needed.
Two details matter. The timezone has to follow the exit, or a Tokyo IP with a New York clock gets flagged. And UDP (HTTP/3, WebRTC) has to leave through the same exit as TCP, or a WebRTC address that differs from the HTTP one gives you away. In a local leak test (Oct 2026, 2 to 5 sessions per setup), ours was 5 of 5 clean, and one third-party SOCKS5 pool leaked in 1 of 3.
Geo-blocks need a new country
Re-rolling IPs inside one country is weak. In an internal eval (412 paired tasks, two models, July 2026), agents changed their exit 44 times after a block and 12 cleared it. Same-country re-rolls cleared 5 of 15. The eval can't prove a country switch beats a re-roll.
The bounty can. Six of 82 sites failed every run from the US and opened from a local proxy. Re-rolling US IPs would never have fixed those. From Oct 16 to 22, sessions asked for 188 different country codes (staff and test included), and 12.2% of managed-proxy sessions used a non-US country, mostly India, Turkey, Germany, Spain and the UK.
The prompt turns this into a ladder: change one thing at a time, read which countries were already tried, try once with no proxy, then a country that fits the content. If the wall holds, stop cycling browsers and find another route to the data. It describes patterns and never names anti-bot products, because product names age out.
Start cheap, escalate on a block
Residential proxies cost real money. On a flight-search task, our fastest agent paid about $0.0095 per run for one, more than its model calls. On five sites a datacenter IP got 10 of 15 tasks right against 11 of 15 with a proxy. Too small to prove anything, but no sign the proxy bought accuracy. So that agent starts on a datacenter IP at $0.20/GB instead of $5/GB and escalates on a block. Google is the exception: on Nov 12 one datacenter workload produced 68% of the day's datacenter hard blocks, all Google refusing the IP. Prices are in our proxy provider comparison, and you can bring your own.
One click beat Akamai
Mouse moves plus one reload do work inside our Akamai handler. It waits, moves the pointer like a person for a few seconds, and reloads once. On production Windows-persona browsers in October it cleared 29 of 30 blocked sessions on 4 retail sites, in 5-7 seconds each. Those were 30 chosen sessions. Across all Akamai encounters the rate is the 24.6% above.
Then there's typing. On a hotel login behind Akamai (Nov 15), filling the form from JavaScript got the login request a 403 in 7 of 8 sessions. Real mouse and keyboard events over CDP: 0 of 8. The default input path of our open-source browser-harness: 0 of 8. On the five other sites in that test, JavaScript input was never blocked. All seven blocks were this one login.
A follow-up pinned it on focus. Every arm that focused the field from JavaScript was blocked (11 of 11). Every arm that clicked it with a real mouse event got through (0 of 11 blocked), even when JavaScript then set the value. It's one site with small arms, so treat it as a lead. We haven't shipped an instruction for it.
What did nothing
We tested two popular theories and looked at a third.
Production traces made deep links look dangerous: 2.25% of deep-link navigations were followed by a block, against 0.37% for other steps. That smelled like "warm up on the homepage first", so we tested it live. No effect. Deep links probably correlate with blocks because they point at the pages sites protect. One DataDome site did flip (deep links blocked 4 of 4, homepage first 0 of 4), so we didn't rule it out, and we didn't ship "visit the homepage first" either.
Hiding Runtime.enable is said to get you detected. It made no difference. For what a page can see, read what browser fingerprinting is.
The third, agents wiggling the mouse themselves, is not a test: calls with 3 or more moves were blocked 0 of 7 times against 5 of 3,099 for the rest. Too few to say anything, so mouse movement stays inside the solvers.
What's still broken
The seven bounty sites are our solver backlog. No IP, country or typing trick gets past a puzzle you can't read. A GeeTest v4 icon solver has merged since, and we haven't re-run that site.
There's no solver in the EU region yet, so every captcha number here comes from US browsers. And 28 sessions per arm can hide a small effect.
If you want to see where your own workload gets stuck, run it on Browser Use Cloud and watch the captchaSolverStarted and captchaSolverFinished events.
Published November 20, 2026
Keep reading
All articlesEngineering
/Nov 20, 2026Two Config Flags Cut Our Browser Agent's Flight Search From 30.7 s to 20.7 s
One Google Flights task went from 30.7 s to about 18 s. The causes were boring, and you can check every one in your own agent today.
Magnus Müller
Engineering
/Nov 20, 2026Our Eval Said +5.1. Up to Half of It Was Our Own Bug.
A fix scored +5.1 on our agent eval. Then we rejudged the saved runs and found our own harness had been feeding the judge a fifth of each answer.
Magnus Müller
Engineering
/Nov 20, 2026We Deleted 137,000 Lines of Our Own Code and Benchmark Scores Went Up
We built a scheduler UI, deleted it, and wrote a 37-line prompt that replaced it. browser-use 1.0 follows the same idea: one main tool, python.
Magnus Müller
AI Trends