Sponsored content: this article was produced in partnership with Bright Data. How thehype labels partner materials is set out in our editorial policy.
An AI agent that browses for you runs into the same CAPTCHA and 403 pages as any bot. Companies have dealt with this in very different ways. OpenAI's agent simply clicked a checkbox. Browserbase gave its browsers a signed ID, and Perplexity and SerpApi ended up in court. Below are five such cases and what each approach costs.
Why agents hit a wall
Websites built their defenses against scrapers long before agents existed. Security systems such as Google's reCAPTCHA were made to keep bots off the internet. A CAPTCHA, a JavaScript challenge and an IP filter cannot tell a shopping assistant from a spam bot. Both arrive without a human at the keyboard.
The walls got higher in 2025. On July 1, Cloudflare began blocking AI crawlers by default on new domains. In January of that year Google had deployed SearchGuard, which sends a JavaScript challenge to unrecognized search queries and blocks automated ones.
So every company that ships an agent has to pick a way through. The first case shows the simplest one.
Case 1. OpenAI: the agent that clicked the checkbox
In July 2025, a Reddit user posted screenshots of ChatGPT agent working through a video conversion task. On the way it met Cloudflare's "Verify you are human" checkbox and clicked it. Tom's Hardware quoted the agent's own narration: "So now I'll click the 'Verify you are human' checkbox to complete this verification."
The story spread as "AI passes the robot test", which overstates it. The agent ticked a checkbox. It did not solve a puzzle of distorted images, and Cloudflare's check looks at the browser and its behavior more than at the click.
What happened a month later matters more. On August 28, 2025, Cloudflare launched "signed agents", and ChatGPT agent was in the first group. A signed agent attaches a cryptographic signature to its requests, so a site knows who is behind them. OpenAI's agent now shows an ID at the gate it once slipped past.
That ID system has a second launch partner whose whole product is a browser for agents.
Case 2. Browserbase: a passport and a solver
Browserbase runs cloud browsers that AI agents control. On August 28, 2025, it announced its work with Cloudflare on Web Bot Auth and called the technology "a passport for your AI agent". The agent signs each request, the site checks the signature and then decides what to allow.
The company is careful about the limits. Its docs say a valid signature "doesn't grant access" to a site or sign the agent into an account. Web Bot Auth is still offered in beta, and it only helps on sites that take part.
For all other sites, Browserbase keeps a second tool. Its docs state that CAPTCHA solving is enabled by default for all sessions and can take up to 30 seconds. The same page says to use it only for workflows that users and website owners authorize.
So Browserbase identifies its agents where a site supports that, and solves the challenge everywhere else. The next case shows what happens when a site decides the agent should not be there at all.
Case 3. Perplexity: accused of hiding, then sued for shopping
Perplexity appears in two separate disputes. The first is about crawling. On August 4, 2025, Cloudflare accused it of running an undeclared crawler. Cloudflare said the declared Perplexity bot made 20–25 million requests a day. Once it was blocked, a second crawler with a generic Chrome user agent sent 3–6 million more. Perplexity said the report was full of technical errors.
The second dispute is about an agent acting for a logged-in user. Amazon sued Perplexity in November 2025 over Comet, its browser that can shop on a user's behalf. In March 2026, Judge Maxine Chesney ordered Perplexity to stop accessing password-protected Amazon accounts, Decrypt reported. Amazon also changed its seller agreement from March 4, 2026, to require AI agents to identify themselves.
The order did not last. On August 4, 2026, a Ninth Circuit panel vacated it, Courthouse News reported. The judges wrote that liability here "would require a novel interpretation far afield" from the purpose of the anti-hacking law. The case went back to the district court.
For now, an agent that acts with its user's permission is hard to ban under that law. Courts have been less sure about the search page.
Case 4. SerpApi: the search page as a locked door
SerpApi sells Google results as an API, which means getting past SearchGuard at scale. Two lawsuits test whether that is legal, and the judges have read the same system differently.
Google sued SerpApi on December 19, 2025, in the Northern District of California. It argued that bypassing SearchGuard breaks the anti-circumvention rule of the DMCA. On July 20, 2026, Judge Yvonne Gonzalez Rogers granted SerpApi's motion to dismiss, PPC Land reported. Claims about results without copyrighted content were dismissed for good. Google got 21 days to refile the rest.
Reddit filed its own suit in New York on October 22, 2025, against SerpApi, Oxylabs, AWMProxy and Perplexity. It says the scrapers took Reddit posts out of Google's results. Three of them, by its count, accessed nearly 3 billion result pages in two weeks of July 2025. In July 2026, Judge Paul Engelmayer let the anti-circumvention claims against Perplexity and SerpApi go forward, MLex reported. He found that SearchGuard does control access to online work.
So the law on CAPTCHA bypass is unsettled, and it depends on what sits behind the check. That uncertainty is one reason many teams don't build this part themselves.
Case 5. Bright Data: unblocking as a service
Bright Data sells the whole problem as infrastructure. Its Web Unlocker API takes a URL and returns the page, with proxies, retries and CAPTCHAs handled on its side. Its MCP server wraps that into 30 tools an agent can call, with 5,000 free requests a month.
One run shows the wall and the way around it. An agent asked to "find what users hate about Cursor on Trustpilot" fetches the page and gets "403 Forbidden".
With the MCP server connected, the agent never opens the page. It finds a ready-made Trustpilot scraper with search_scrapers, starts it with run_scraper and reads the rows from get_scraper_results. The run returned 209 reviews in about 3 minutes, and 56% of them mention billing and refunds.
The trade-off is the one from Case 4. The agent gets through because it does not announce itself. Bright Data's position is that public pages are open to read.
The last section compares the five approaches.
Which way through fits which agent
The approaches differ in who has to agree. The table below sets them side by side.
| Approach | Case | Works when | Weak point |
|---|---|---|---|
| Click through | ChatGPT agent | the check is a simple checkbox | fails on real puzzles |
| Signed identity | Browserbase, ChatGPT agent | the site supports Web Bot Auth | still a beta, few sites |
| Acting for a logged-in user | Perplexity Comet | the user gives permission | the site may sue |
| Getting past search defenses | SerpApi | the data is in search results | two courts disagree |
| Unblocking service | Bright Data | the page is public | cost per request |
So how should an AI agent deal with CAPTCHAs and bot blocks? It depends on what the agent does. An assistant that acts for one user inside their accounts should identify itself, and the Ninth Circuit ruling gives it some cover. An agent that reads public pages at volume needs infrastructure built for that, and the Bright Data run shows how little code it takes. Both routes are still open questions. Signed identity covers few sites so far, and the courts have not agreed on where reading ends and circumvention begins.
FAQ
Short answers to the questions these cases raise most.
Can AI agents solve CAPTCHAs?
Some can click a simple checkbox, as ChatGPT agent did in July 2025. Services such as Browserbase solve harder challenges for the agent, which can take up to 30 seconds.
What is Web Bot Auth?
It is a way for a bot or agent to sign its HTTP requests cryptographically. A site that supports it can verify who sent the request and set rules for that agent.
Is it legal for an agent to bypass a bot check?
It is unsettled. In 2026 one federal judge dismissed Google's claims against SerpApi, and another let Reddit's similar claims go forward.
What does Bright Data's Web Unlocker do?
It takes a URL and returns the page. Proxy choice, retries and CAPTCHA handling happen on Bright Data's side.
Is Bright Data MCP free?
Every account gets 5,000 free requests a month, with no credit card required. Beyond that, its README lists $1.50 per 1,000 results.
Sources
Figures in the Bright Data case come from a demo run in November 2026.
- The age of agents: cryptographically recognizing agent traffic – Cloudflare, August 28, 2025
- Cloudflare & Browserbase: pioneering identity for AI agents – Browserbase, August 28, 2025
- CAPTCHA solving, Browserbase docs
- Google loses DMCA bid to treat search scraping like DVD piracy – PPC Land, July 21, 2026
- Reddit sues data scrapers and Perplexity over unauthorized content access – PPC Land, October 22, 2025
- Bright Data MCP server, GitHub – tools, free tier and pay-as-you-go prices
- Bright Data MCP server
- Web Unlocker, Bright Data