TL;DR: A browser agent gets blocked for one of four reasons: a captcha, a banned IP, a sensor script that flagged its behavior, or a country wall. Each needs a different fix, and the wrong fix wastes the run. In our $100 bounty, 7 of 82 testable sites blocked our agent every time, all because of captchas our solver didn't support yet, and 6 more opened once we used a proxy in their own country. On one Akamai login, focusing a field from JavaScript got blocked 11 of 11 times, while a real mouse click got through every time.
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 blocked us only because we came from the US, and opened from a proxy in their own country.
Our fingerprint is good (see our bot detection and stealth benchmark posts), and it still hits walls. What matters is what the agent does next.
Four reasons a browser agent gets blocked
Most blocks are one of four things. A puzzle, which you solve and move on. A banned IP, where solving does nothing and you need a new exit. A sensor flag, where a script on the page didn't like how you behaved, so a new IP does nothing and you have to behave differently. And a country wall, where 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.
Captchas: which anti-bot vendors a solver can beat
From November 8 to 15, our US cloud browsers met 1,394,724 captchas, and 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, but a solver can't fix either of them. Most DataDome misses are pages closed mid-solve, and without those it's 82.3%. Akamai is mostly a 403 that our driver reloads once. If the 403 survives the reload, the IP is banned or a sensor flagged us, and that's a different problem.
A solver also has to know when to quit. Google's "unusual traffic" page shows a reCAPTCHA, but solving it doesn't lift the block, because the block is on the IP. 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. Before that, one site cost us 106 solver starts and 106 failures in an hour.
The agent also has to wait for the solver. In July our older agent capped its wait at 3 seconds and gave up after one wait and one reload. On 43 sign-up forms it ended captcha-blocked on 17, and the newer agent, which waits, on 10. Our default agents now get this line:
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 busy (why that matters). It gets a solve_captcha tool instead, which turns the solver on, waits up to 20 seconds for a captcha to show up and up to 95 seconds for it to finish, then turns it off again. On staging it called the tool unprompted and solved hCaptcha 6 of 6 times, in 21 to 30 seconds each. The same advice as a paragraph of prose went 0 of 3. Three tries is nothing, but I'd still give an agent a tool before a paragraph.
On Stealth Bench V2 our anti-detect browser got through on 90.5% of 200 visits to 100 hard sites with the solver on, and 84.0% with it off. Chrome with a US proxy got 14%. We picked those sites because our browser had got through and stock Chrome hadn't, so the comparison flatters us. More in captcha solving for AI agents.
Banned IP: change it without losing the login
A fresh browser fixes a burned IP, but it also 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 November 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 if you build your own. The timezone has to follow the exit, or a Tokyo IP with a New York clock gets flagged. And UDP traffic (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 small local leak test in October, ours was clean in 5 of 5 sessions, and one third-party SOCKS5 pool leaked in 1 of 3.
Geo-blocks: change the country, not the IP
Re-rolling IPs inside one country is weak. In an internal eval in July (412 paired tasks, two models), agents changed their exit 44 times after a block and 12 of those cleared it. Same-country re-rolls cleared 5 of 15.
The bounty made the case more clearly. Six of the 82 sites failed every run from the US and opened from a local proxy, and no amount of re-rolling US IPs would have fixed those. Our prompt turns this into a ladder: change one thing at a time, check which countries were already tried, try once with no proxy, then pick a country that fits the content. If the wall holds, stop cycling browsers and look for another route to the data.
Sensor flags: how the agent types matters
Mouse movement and 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 to 7 seconds each. Those were 30 chosen sessions, though. Across all Akamai encounters the rate is the 24.6% above.
Typing turned out to matter more than we expected. On a hotel login behind Akamai on November 15, filling the form from JavaScript got the login request a 403 in 7 of 8 sessions. Real mouse and keyboard events over CDP were blocked 0 of 8 times, and so was the default input path of our open-source browser-harness. On the five other sites in that test, JavaScript input was never blocked.
A follow-up pinned it on focus. Every arm that focused the field from JavaScript was blocked (11 of 11), and every arm that clicked it with a real mouse event got through (0 of 11 blocked), even when JavaScript set the value afterwards. It's one site with small arms, so treat it as a lead, not a rule. We haven't shipped an instruction for it yet.
Anti-bot tricks that did nothing
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 suggests "warm up on the homepage first", so we tested it live on 7 protected sites. Deep links were blocked 4 of 28 times and homepage-first 3 of 28 (p = 1.0). Deep links probably correlate with blocks because they point at the pages sites protect most. One DataDome site did flip (4 of 4 against 0 of 4), so we haven't ruled it out, but we didn't ship it either.
Hiding Runtime.enable is supposed to keep you from being detected. On 6 sites it made no difference: 4 of 67 blocked with it on, 3 of 68 with it off.
What still blocks our agent
The seven bounty sites are our solver backlog, and no IP, country or typing trick gets past a puzzle you can't read. A GeeTest v4 icon solver has merged since, but we haven't re-run that site. There's no solver in our EU region yet, so every captcha number here comes from US browsers.
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 articles
Engineering
/Nov 20, 2026browser-use 1.0: Why We Replaced 24 Agent Tools With One
browser-use 1.0 gives the model one main tool, python, instead of 24 browser actions. The same models scored 14 to 33 points higher, and the agent started doing things we never built.
Magnus Müller

Engineering
/Nov 20, 2026How to Make a Browser Agent Faster: Six Causes We Found
One Google Flights task took our browser agent 30.7 seconds. Six causes explained most of it, and two config flags alone brought it to 20.7 seconds.
Magnus Müller

Engineering
/Nov 20, 2026LLM-as-a-Judge Bug: Our Judge Read a Fifth of Each Answer
A fix scored +5.1 points on our agent eval. Up to half of that came from a bug: our LLM judge only saw the end of each answer. How rejudging saved runs caught it.
Magnus Müller
AI Trends