What Is a Captcha Solver? How Bypassing Works and How to Defend

Captchas guard almost every login, sign-up and checkout form. At the same time, there is an entire industry that knows how to get past them: captcha solving services, open-source tools, trained neural networks. In this article we explain what a captcha solver is, who uses one and why, where the line between automation and abuse runs, and most importantly, why a captcha alone does not protect a website and what you need to add to make protection work. The article is written for developers who build protection into their projects, site owners and bloggers dealing with spam and fake sign-ups, and anyone who wants to understand what happens behind the "I'm not a robot" checkbox.

What a captcha is and why it exists#

CAPTCHA stands for Completely Automated Public Turing test to tell Computers and Humans Apart. The idea is simple: give a task that is easy for a human and hard for a program. Whoever solves it is most likely a human.

A brief history#

• 1997. The AltaVista search engine was struggling with bots that mass-submitted URLs to its index. Its engineers came up with distorted text that had to be retyped: the OCR software of the time could not read it. This is considered one of the first captcha prototypes.
• 2000-2003. Researchers at Carnegie Mellon University (Luis von Ahn, Manuel Blum, Nicholas Hopper and John Langford) formalized the concept and gave it the name CAPTCHA. One of the first practical uses was protecting Yahoo chat rooms, which bots were flooding with advertising links.
• 2007. reCAPTCHA appeared. Its twist was that, along with the control word, users transcribed a word from digitized books and newspaper archives that a machine had failed to read. Millions of people helped digitize texts for free. Google acquired the project in 2009.
• 2014. Google published research showing that its own neural network read the hardest variants of distorted text with about 99.8% accuracy. Text captchas stopped being reliable, and they were replaced by reCAPTCHA v2 with the "I'm not a robot" checkbox and image grids ("select all images with traffic lights"). The checkbox evaluated not just the click, but also how the browser behaved before it.
• 2018. reCAPTCHA v3 was released. It shows the user nothing at all: it assigns a score from 0.0 to 1.0 in the background, and the website makes the decision. hCaptcha appeared around the same time, and in 2020 Cloudflare switched to it.
• 2022 and beyond. Cloudflare launched Turnstile, a captcha that in most cases checks the browser without any challenge. Apple, together with Cloudflare and Fastly, introduced Private Access Tokens: the device itself attests that a real user is behind it. Meanwhile, sliders and puzzles (GeeTest, Tencent, Yandex SmartCaptcha) spread widely: the user drags a piece of an image into place.

The trend is clear: captchas are moving away from "a task for a human" toward "an invisible risk assessment". The reason is that machines have learned to solve almost any visual task. In 2023, researchers at the University of California, Irvine compared humans and bots on popular captchas, and the bots turned out to be not only faster but, in many cases, more accurate than humans. In 2024, a team at ETH Zurich showed that a computer vision model, according to the authors, passes reCAPTCHA v2 image challenges in 100% of cases.

Why websites add a captcha#

A captcha is useful wherever a single automated action costs an attacker nothing, while a million such actions cause real damage. Typical places:

• login forms (protection against password guessing);
• sign-up (protection against mass account creation);
• password recovery and SMS sending (every SMS costs money);
• comments, reviews, contact forms (protection against spam);
• checkout, promo codes, polls;
• search and heavy pages that load the server.

What a captcha protects against#

Threat	What it looks like	Consequences
Brute force and credential stuffing	A bot tries login/password pairs from leaked databases	Account takeover, theft of money and data
Mass sign-ups	Thousands of fake accounts	Abuse of bonuses and free trials, spam, fake engagement
Spam	Automated comments and messages	Damaged reputation, search engine penalties
SMS and email bombing	A bot submits a "send me a code" form for other people's numbers	Direct SMS costs, mail domain blocklisting
Scalping	A bot buys tickets or scarce goods within seconds	Lost customers, resale at inflated prices
Aggressive scraping	High-frequency collection of a catalog or prices	Server load, loss of unique content
Fake engagement	Fake votes, clicks, views	Distorted analytics, ad fraud

It is important to understand that a captcha does not make these attacks impossible. It makes every attempt more expensive and slower. Where a bot used to submit a form in 50 milliseconds for free, with a captcha it now needs time and money to solve it. For many attacks, that is enough to make them unprofitable. For others it is not, and we will get to that below.

Who bypasses captchas and why#

A common stereotype says that only hackers and spammers bypass captchas. In practice, many of the people who automate captcha solving are entirely legitimate users. It helps to split the scenarios into three groups.

White-hat scenarios#

• QA engineers and automated tests. In a test environment, the captcha can often be disabled or switched to test keys. But end-to-end tests in production, checks of the real captcha integration, or tests of third-party widgets require the captcha to be passed automatically.
• Monitoring your own services. Synthetic monitoring regularly walks through a user journey (login, checkout) and reports if something breaks. A captcha along that path should not get in the way of the check.
• Accessibility. For people with visual or motor impairments, a captcha can block access to a service entirely. Assistive tools help them pass the check.
• Automating your own workflows. Exporting reports from a bank or supplier portal where the company has an account but no API.
• Research. Academic and journalistic research, archiving publicly available information, security testing under a contract with the system owner (penetration testing).

The gray zone#

• Collecting public data. Competitor price monitoring, job and listing aggregators, SEO analytics. The data itself is public, but the website's terms may prohibit automated collection.
• Managing many accounts. SMM specialists and agencies that run dozens of client accounts. Legal as long as it does not break the platform's rules, and many platforms restrict it.
• Automated booking and buying. A bot that grabs a free doctor's appointment or a visa center slot. For one person it is a convenience; at scale it turns into reselling slots.

In the gray zone, legality depends on the details: the rules of the specific resource, the volume of requests, what data is collected and what happens to it afterward.

Black-hat scenarios#

• password guessing and taking over other people's accounts;
• mass sign-ups for bonuses, spam and fraud;
• sending spam through forms;
• scalping tickets and scarce goods;
• fake votes, reviews and ad clicks;
• theft of personal data and content.

Bypassing a captcha is not a violation in itself#

Technically, a captcha is just an interface element. The laws of most countries contain no specific ban on solving one automatically. What makes it a violation is the purpose of the bypass and the specific rules it breaks:

• the law: unauthorized access to computer systems, fraud, violations of personal data laws, spam;
• the resource's terms of use: many websites explicitly prohibit automated access in their terms, and breaking them can lead to a ban or a civil lawsuit.

If both conditions are met, meaning local law does not prohibit what you are doing and the resource's rules allow it, passing a captcha automatically does not violate anything in itself. We cover the legal side in more detail in a separate section.

What is a captcha solver#

A captcha solver is any tool or service that automatically passes a captcha check on behalf of a human and returns the result: recognized text, click coordinates, or a token confirming the check was passed. The term covers things that work very differently, from a cloud API to a neural network on a developer's laptop. Let's go through the main options.

First: what "solving a captcha" actually means#

Captchas come in two fundamentally different types, and that determines how a solver works.

Image captchas. The user is shown an image: distorted text, a grid of pictures, a puzzle. The solution is the answer to the task: text, image numbers, click coordinates or a slider offset. The solver only needs the image.

Token captchas. reCAPTCHA, hCaptcha, Cloudflare Turnstile and Yandex SmartCaptcha work differently. The widget on the page checks the browser and the user, then issues a token, a long string that is inserted into the form. The website's server sends this token to the captcha provider for verification and gets an answer: did the user pass the check or not. Here the solver needs a valid token for a specific website. To get one, it uses the site's public key (sitekey) and the page URL, and sometimes a proxy so that the check comes from the right IP.

A token has a limited lifetime: about two minutes for reCAPTCHA and about five for Turnstile, and it can be used only once.

Captcha solving services with an API#

The most common option. This is a cloud service: you send it a task over an HTTP API and pick up the finished solution a little later. The flow is almost the same everywhere:

1. The client submits a task with the createTask method: the captcha type and its parameters (an image, or the page URL and sitekey).
2. The service returns a task ID, taskId.
3. Every few seconds, the client requests the result with the getTaskResult method.
4. When the solution is ready, the service returns it, and the client puts the answer into the form or request.

Inside the service, tasks are solved by:

• humans. Historically the first option: a distributed network of workers who solve the challenges shown to them for a small fee. Slower than machines, but they can handle new and unusual captcha types;
• neural networks. Computer vision models for images and automated browsers for token captchas. Fast and cheap for mainstream types;
• a hybrid setup. Most modern services combine both approaches: a machine handles typical tasks, while hard and new ones go to humans.

Pros for the client: nothing to train or maintain, pay per solution, dozens of captcha types through a single API. Cons: every solution costs money, there is a delay of a few seconds, and task data (the page URL, the image) is shared with a third party.

Local tools#

A solver does not have to live in the cloud. Some tools run on the user's computer or the developer's server:

• OCR libraries (for example, Tesseract) with image preprocessing: denoising, deskewing, binarization. They still handle simple homemade text captchas.
• Browser extensions. For example, extensions that pass the audio version of reCAPTCHA using speech recognition. They are often created as accessibility tools in the first place.
• Plugins for automation frameworks (Playwright, Puppeteer, Selenium). They hide the signs of an automated browser and either solve the captcha locally or hand it off to a cloud service.
• Computer vision algorithms for sliders. For many puzzles, it is enough to find the outline of the cut-out piece (for example, with OpenCV), calculate the offset, and then generate a "human-like" mouse trajectory.

Local tools are free and do not send data anywhere, but they require setup and break as soon as the captcha provider changes something.

Custom solutions and trained models#

Companies that need to pass captchas at high volume sometimes build a solver themselves:

• Classifiers and detectors. A convolutional neural network for reading a specific captcha's text is trained on a few thousand labeled examples. For image grids, object detectors such as YOLO are used.
• Behavior models. Generating mouse movements, typing speed and pauses that look human to behavioral analysis.
• Browser farms. A pool of real browsers with different profiles and proxies that pass token captchas the same way a user would.

This path is expensive: it requires training data, infrastructure and constant maintenance. It only pays off at very high volumes or for an exotic captcha that services do not support.

Multimodal AI models#

A separate class that has emerged recently. Language models that can work with images solve many visual challenges without any special training: they describe what is in a picture, read distorted text, find the right objects. AI agents that control a browser go through entire user journeys, captcha included. Model providers usually restrict this kind of use through their policies and built-in refusals, but open models have no such safeguards. For defenders, this means the boundary of "what only a human can do" is moving even faster.

Comparing the approaches#

Approach	What you need	Speed	Cost	Resilience to captcha changes
API service	An API key and a couple of HTTP requests	Seconds	Pay per solution	High: the service handles support for new versions
Local tool	Installation and setup	Fast	Free	Low: breaks on updates
Custom model	Data, training, infrastructure	Fast	High upfront cost	Medium: needs retraining
Multimodal AI model	Access to a model	Seconds	Pay per request or your own hardware	Medium: depends on the challenge type

So the short answer to "what is a captcha solver" is this: anything that automatically turns a captcha challenge into an answer the website accepts. A human on the other side of an API, an OCR script, a browser extension, a trained neural network, or a combination of these.

Is bypassing a captcha legal#

What determines legality#

There is no single "captcha law"; legality is made up of several layers:

1. Computer misuse laws. Almost every country has laws against unauthorized access to computer systems: the Computer Fraud and Abuse Act (CFAA) in the US, the Computer Misuse Act in the UK, and equivalents elsewhere. If a captcha bypass is used to reach something you have no right to access (other people's accounts, restricted areas), it is no longer automation.
2. Personal data. The GDPR in the EU, the CCPA in California and similar laws regulate the collection of personal data regardless of whether it is publicly visible.
3. Rights to content and databases. Texts, photos and the website's database itself may be protected by copyright or, in the EU, by the database right under the Database Directive.
4. The resource's terms of use. A user agreement is a contract. Breaking it is usually not a crime, but it can be grounds for a ban and a civil lawsuit.

Two US cases show how nuanced this is. In the long-running hiQ Labs v. LinkedIn dispute, courts concluded that collecting public profiles does not violate the CFAA in itself. Yet in 2022 the court found that hiQ had breached LinkedIn's user agreement, and the case ended in a settlement under which hiQ agreed to stop scraping and delete the data it had collected. In Meta v. Bright Data (2024), the court sided with the scraper: Bright Data collected public data without logging in, and Meta's terms were found not to cover that. Meta later dropped the case.

One more important point: what matters is not only the law of your own country, but also the law of the country where the target resource and its users are located.

Our position at captcha-solver#

We build a service for lawful automation and explicitly prohibit malicious use. This is set out in two documents that every customer accepts:

• Terms of Service: the service may be used only for lawful purposes; it is prohibited to bypass the protection of systems you do not own and have no explicit permission to test, to send spam, or to engage in any activity that breaks the law.
• Acceptable Use Policy: permitted scenarios (testing your own applications, CI/CD, accessibility tools, data collection that complies with the target site's terms and the laws of the country where it is located) and prohibited ones (credential stuffing and brute force, bypassing the protection of third-party websites in violation of their terms, mass account creation, spam, fake engagement, fraud).

We monitor API usage for abuse and suspend accounts that break these rules. Responsibility for complying with the law and the target resource's terms lies with the customer, and not knowing the applicable law is no excuse.

How to make captcha bypass harder on your website#

Now let's look at the situation from the defender's side. The main idea of this section: a captcha is just one layer of many, and the more independent layers you have, the more expensive an attack becomes. The measures below go from simple to advanced.

Start with a correct captcha integration#

A surprising number of websites are "protected" by a captcha that can be bypassed without solving anything at all. Check your integration:

• The token is verified on the server. If the check only exists in JavaScript, a bot simply submits the form directly. The server must call the provider's verification method (siteverify for reCAPTCHA, hCaptcha and Turnstile) on every form submission.
• Response fields are checked, not just success. Verify hostname (the token was solved on your domain, not on someone else's site with the same key), action if the provider returns it (the token was issued for this form), the issue time, and for reCAPTCHA v3, the score threshold.
• The token is single-use. The provider will not accept a token twice, but that only helps if you actually send it for verification every time instead of caching the result per session.
• There is no fallback path without a captcha. A mobile app, an old API version, a separate partner endpoint: bots quickly find the door with no captcha.
• A provider outage does not open the door. Decide in advance what happens if the captcha service is unavailable: let everyone through (fail open) or no one (fail closed). Failing open on a login form is dangerous.

Rate limit by several keys#

A limit of "N requests per minute per IP" breaks with the first proxy pool. You need limits by several independent keys:

• by IP and subnet;
• by the account or email being targeted by login attempts, regardless of IP: that way, rotating proxies does not extend a brute-force attack;
• by device or browser fingerprint;
• a global limit per action (for example, no more than a certain number of SMS per hour across the whole service).

Increment the counter atomically and before running the check itself, otherwise a burst of parallel requests slips past the limit. Use windows of different lengths: a short one cuts bursts, a long one catches slow brute force designed to wait out the short window.

Analyze where traffic comes from#

• Network type (ASN). Sign-up requests from hosting and cloud provider ranges are more suspicious than those from residential ISP networks.
• IP reputation. Commercial and open databases flag proxies, VPNs, Tor exit nodes and addresses seen in attacks. You can build your own IP rating on top of them: every suspicious episode lowers the score, long normal activity raises it.
• Consistency. The browser's time zone, language, IP geolocation and account region should roughly match. A browser with a Tokyo time zone, a German interface and an IP from Brazil is not guilty in itself, but it is worth a closer look.
• Technical fingerprints. TLS handshake fingerprints (JA3/JA4), HTTP/2 parameters and header order give away an HTTP client that claims to be Chrome but is not.

Collect behavioral signals#

• Form completion time. A human does not fill in a sign-up form in a fraction of a second.
• Honeypot field. A field hidden from humans that bots fill in automatically.
• Navigation path. A user first opens the page, then submits the form. A bot often sends a POST right away, with no prior GET and without loading styles and scripts.
• Interaction with the interface. Mouse movements, field focus, typing speed, scrolling. These signals work well in combination, but each can be faked on its own.

Tripwires and trigger scenarios#

A tripwire is an action that cannot be performed through the normal interface. If it happens, you are dealing with a script or a person who is tampering with requests by hand. Examples:

• a request contains fields or data types that the form never sends;
• a value has changed that the interface does not allow users to change;
• a burst of identical parallel requests within milliseconds;
• a request contains tampered values that the server is supposed to calculate (a price, a discount, someone else's ID);
• requests to paths that do not and cannot exist on the website (typical vulnerability scanner paths).

A good practice is to respond to a tripwire not with an instant ban, but with a "shadow" mode and a delayed block. If the ban comes immediately, the attacker can easily tell which action gave them away and work around that check. If the account keeps "working" but can actually do nothing, and the block comes later, figuring out the cause is much harder.

Business logic as protection#

Often the most effective measure is not technical at all:

• one mailbox, one account: keep in mind that with many email providers, a single mailbox accepts mail for several variants of an address;
• blocking disposable email domains at sign-up;
• email or phone verification before access to valuable actions;
• bonuses and free trials granted after a confirmed action (payment, linking a card), not right after sign-up;
• delayed payouts and withdrawals for new accounts.

These measures do not get in the way of real users, but they break the economics of mass sign-ups.

Combine everything into a risk score#

Individual signals are noisy, so decisions are best made on a combined risk score:

Risk score	Action
Low	Let through without a captcha or with an invisible check
Medium	Show a captcha, add a delay, ask for email confirmation
High	Block, switch to shadow mode, raise an alert

This adaptive approach is better than showing a captcha to everyone all the time: honest users barely see any checks, while suspicious traffic gets more obstacles.

And monitoring is a must. Log captcha failures, rate limit hits and tripwire triggers, and set up alerts on spikes. An attack you learn about a week later from customer complaints costs far more than one you spot within five minutes.

Protection is built around the business#

There is no universal recipe. An online store needs to protect promo codes and stock, a bank needs to protect logins and transfers, a media site its comments, a SaaS its free trials. Start with the question "which automated action costs us the most money" and protect that first. Analyze real attacks in your logs: they will show which layers you actually need.

Why a captcha alone is not enough#

A captcha checks a moment, not an intent#

A captcha answers a narrow question: "did someone pass the check just now". It does not know who that was: a human, a neural network, or a worker at a captcha solving service paid a fraction of a cent. It does not see that this is the thousandth attempt in an hour to log into someone else's account. That is why access decisions should be based on context, with the captcha being just one of the signals.

Advanced solutions exist, but they are expensive#

There are bot protection systems on the market that show very high effectiveness. A good example is DataDome: it is not just a captcha but a platform that analyzes every request in real time across many signals (device and network fingerprints, behavior, reputation, machine learning on traffic from thousands of websites) and shows a challenge only when the risk is genuinely high. Akamai Bot Manager, HUMAN Security, Kasada and Cloudflare's enterprise Bot Management belong to the same class.

The problem is the price. These solutions target large enterprises, and a mid-sized project usually cannot afford them: the subscription may cost more than the bot damage it prevents. That is why most websites live with a free or inexpensive captcha plus the in-house measures described above.

Anything can be bypassed; the question is the cost#

Even the most advanced protection can be bypassed. The right question is not "can it be done" but "how much does it cost and does it pay off". Every protection layer raises the cost of a single successful attempt: solving the captcha, a quality proxy, a fresh browser profile, time spent imitating behavior. An attack makes sense as long as the expected gain exceeds the cost.

The practical takeaway: your protection does not have to be impenetrable. It has to make attacking your website more expensive than the gain from it, and more expensive than attacking the website next door.

What a modern bot looks like#

Against serious protection, attackers no longer use a simple requests script but a whole toolkit:

• anti-detect browsers with hundreds of profiles, each with its own plausible fingerprint: graphics card, fonts, screen resolution, time zone, language;
• residential and mobile proxies, meaning traffic routed through the IP addresses of real home users and mobile carriers, which cannot be blocked without hurting real customers;
• human imitation: pauses, curved mouse movements, typos while typing, warming up accounts with weeks of normal activity;
• captcha solving services for the checks that still appear.

To an ordinary captcha, such a bot looks like an ordinary user. It can only be detected either by advanced commercial systems or by your own anti-fraud system: accumulated user behavior history, links between accounts (shared devices, payment details, similar patterns), and anomalies in business metrics.

A game of cat and mouse#

The fight between bots and defenders is an endless race. Defenders release a new check, and within a few weeks a way to pass it appears. Bots adopt a new disguise technique, and defenders learn to spot it. Both sides use machine learning, and both sides keep adapting. Any "final" solution works only until attackers find it worthwhile to take it apart.

Why successful defenses stay quiet#

Defenders have a structural disadvantage: information about truly effective detection methods is almost never published. As soon as it becomes known which signal gives bots away, bot operators remove that signal. That is why vendors keep their signals secret, companies do not talk about their anti-fraud rules, and what remains in the public domain is mostly general principles and outdated techniques.

Attackers, on the other hand, actively share tools and experience on forums and in chat groups. For a defender, this means one simple thing: do not wait for a ready-made recipe from the internet, study your own traffic. Your logs know more about the bots on your website than any article.

Conclusions#

Captchas have evolved from distorted text into invisible risk assessment, but their job has not changed: to make automated actions more expensive. The key points of this article:

1. A captcha solver is any tool that passes a captcha automatically: a cloud service with humans or neural networks, a local script, a browser extension, a custom trained model or a multimodal AI model.
2. Bypassing a captcha is not a violation in itself. Legality is determined by the purpose, the laws of your country and of the target resource's country, and the resource's own terms. Testing, monitoring and accessibility are legitimate scenarios; account takeover, spam and fake engagement are not.
3. A captcha raises the cost of an attack but does not stop it. Any captcha can be passed; the only question is whether it pays off for the attacker.
4. Protection is built in layers. Correct server-side token verification, rate limits by several keys, traffic origin analysis, behavioral signals, tripwires and business logic together achieve far more than any captcha on its own.
5. There is no silver bullet. Commercial anti-bot platforms are effective but expensive, and the fight against bots never ends. The best source of knowledge about the bots on your website is your own logs.

What to be ready for#

Developers. Your captcha will be passed: by a human, a service or a neural network. Design your protection with that in mind. Verify tokens on the server with all fields, set rate limits by several keys, log and monitor. A captcha is one of the walls, not the whole castle.

Site owners and bloggers. For a small website, a modern captcha (Turnstile, reCAPTCHA or SmartCaptcha), a comment spam filter, email verification and blocking disposable addresses are usually enough. If bot traffic grows, first look at your logs to see what exactly the bots are doing, and close that specific scenario.

Everyday users. If you see captchas too often, the cause is most likely your IP's reputation: a VPN, public Wi-Fi or a mobile connection shared by many people. The captcha is not accusing you; it simply does not see enough signals that you are human.

Everyone. The line between humans and machines is blurring, and captchas will become less and less visible while protection becomes more and more behavioral. That means more analysis of user data, so privacy questions will only get louder.

Frequently asked questions#

What is a captcha solver in simple terms?#

It is a program or service that passes a captcha instead of a human. You give it the task (an image or the URL of a page with a captcha) and get back an answer the website accepts: recognized text, click coordinates or a token. The task itself may be solved by humans, neural networks, or both.

Is it legal to use a captcha solving service?#

In itself, yes, as long as your scenario does not break the laws of your country or the target resource's country, or that resource's terms of use. Testing your own applications, monitoring and accessibility tools are legitimate scenarios. Taking over other people's accounts, spam and mass sign-ups are prohibited both by law and by the rules of captcha solving services.

Can a website tell that a captcha was solved by a service?#

From the captcha alone, usually not: the token obtained by the service is valid. But the website sees other signals: where the request came from, how the browser behaved, how many attempts there were. It is those signals, not the captcha, that reveal automation.

Which captcha is enough for a small website?#

No captcha is invulnerable, so what matters more than the specific captcha is what surrounds it. For a small website, a modern free captcha (Turnstile, reCAPTCHA, SmartCaptcha) is usually enough on actions whose abuse costs money: login, sign-up, sending SMS, orders, comments. There is no need to put it on every form: a captcha hurts usability and lowers conversion. Add email verification and rate limiting, and keep an eye on your logs.