Browser Agents Get Blocked Four Ways — and Each Needs a Different Fix

A $100 bounty found 7 of 82 sites blocked a browser agent every time — all unsupported captchas. The fix depends on which of four walls you hit, and the wrong one wastes the run.

Wire notes

  • 7 of 82 bounty sites blocked the agent on every run; all used unsupported captchas, and 6 more opened via local-country proxies
  • 1,394,724 captchas met Nov 8-15; 79.2% solved automatically
  • Real mouse clicks passed an Akamai login 11 of 11 times; JavaScript focus was blocked 11 of 11 times
  • Anti-detect browser cleared 90.5% of Stealth Bench V2 visits with solver on; stock Chrome on a US proxy got 14%
  • IP-swap tool shipped November 8; agents call it ~10 times per 100 runs
Browser Agent Blocked? How to Tell Which Wall You Hit and Get Past It
PhotoBrowser Agent Blocked? How to Tell Which Wall You Hit and Get Past It — StartupStockPhotos / pixabay

By James Calloway

A $100 bounty to find sites that block a browser agent produced 82 testable sites — and seven of them blocked the agent on every single run, all because of captchas the agent's solver didn't support at the time. Six more sites blocked it only because the traffic came from the US, and opened once a proxy in their own country was used.

That's the core finding from Browser Use's engineering team, who laid out the numbers in a technical post this week. The broader lesson: a blocked browser agent hits one of four walls, and each wall demands a different fix. Apply the wrong one and you burn the run.

Why do agents get blocked at all?

The post breaks blocks into four types. A captcha is a puzzle — solve it and move on. A banned IP means solving does nothing; you need a new exit. A sensor flag means a script on the page disliked the agent's behavior, so a new IP won't help. And a country wall means the site serves one country and refuses the rest.

Because these agents work inside logged-in user accounts, every fix has to keep the session alive.

How well do captcha solvers actually work?

From November 8 to 15, the team's US cloud browsers met 1,394,724 captchas, and 79.2% were solved automatically. DataDome and Akamai look weak on paper, but the authors note a solver can't fix either: most DataDome misses are pages closed mid-solve (excluding those, it's 82.3%), and Akamai's 403 usually points to a banned IP or a sensor flag — different problems entirely.

Knowing when to quit matters too. Google's "unusual traffic" page shows a reCAPTCHA, but solving it doesn't lift an IP-level block. In 16 hours of July traffic, 73% of sessions that produced a valid token got challenged again. Before the solver learned to spot that page and rotate the IP, one site cost 106 solver starts and 106 failures in a single hour.

Patience pays. An older July-era agent capped its captcha wait at 3 seconds and ended captcha-blocked on 17 of 43 sign-up forms; the newer agent, which waits, on 10.

Tool beats prose. The fastest agent carries a solve_captcha tool rather than an always-on solver. On staging it called the tool unprompted and solved hCaptcha 6 of 6 times, in 21 to 30 seconds each. The same advice written as a paragraph went 0 for 3.

On Stealth Bench V2, the anti-detect browser got through 90.5% of 200 visits to 100 hard sites with the solver on, and 84.0% with it off. Stock Chrome with a US proxy managed 14%.

How do you swap a banned IP without killing the login?

The team shipped a tool on November 8 that swaps the exit under a running browser instead of restarting it. Agents call it roughly 10 times per 100 runs. Two build details matter: the timezone must follow the exit — a Tokyo IP with a New York clock gets flagged — and UDP traffic (HTTP/3, WebRTC) must leave through the same exit as TCP. In a small October leak test, their setup was clean in 5 of 5 sessions; a third-party SOCKS5 pool leaked in 1 of 3.

Geo-blocks need a country change, not more IPs. In a July internal eval across 412 paired tasks, agents changed exits 44 times after a block and 12 changes cleared it; same-country re-rolls cleared only 5 of 15.

Does how an agent types really matter?

Yes. On an Akamai-protected hotel login tested November 15, filling the form from JavaScript drew a 403 in 7 of 8 sessions; real mouse and keyboard events over CDP were blocked 0 of 8 times. A follow-up pinned it on focus: JavaScript focus was blocked 11 of 11 times, real mouse clicks 0 of 11. The team calls it a lead, not a rule — one site, small samples.

Meanwhile, two popular anti-bot tricks did nothing. Homepage-first "warm-up" made no significant difference on 7 protected sites (4 of 28 blocked on deep links vs 3 of 28, p = 1.0), and hiding Runtime.enable changed nothing across 6 sites.

The seven bounty sites remain on the solver backlog — no IP, country or typing trick gets past a puzzle the machine can't read.

via browser-use.com (Source)

Filed under

Share this article:

More from James Calloway

James Calloway

Show full bio

Market editor covering technology and product news in marketing, advertising, marketing technology, media, SEO and search at Marketing Herald.

401 articles

More on the wire

« Previous article