A CAPTCHA is a speed bump for automation, not a wall. It raises the cost of sending thousands of login attempts, sign-ups or form posts, and it does nothing about flaws an attacker can reach without automation. This guide covers what CAPTCHA stops, how weak deployments get bypassed, how to evaluate a CAPTCHA product on privacy, accessibility and vendor risk, and where it fits next to rate limiting, MFA and breached-password checks.
What CAPTCHA stops, and what it does not
OWASP’s Automated Threats to Web Applications project names the attacks a CAPTCHA is meant to slow down. Four of them map directly to login, sign-up and checkout flows:
- Credential stuffing (OAT-008): username and password pairs stolen elsewhere are tried against your login to find people who reused them (OWASP OAT-008).
- Fake sign-ups (OAT-019, Account Creation): bulk accounts created through normal registration, then used for spam, fraud or fake reviews (OWASP OAT-019).
- Scraping (OAT-011): automated collection of data, sometimes through fake or compromised accounts (OWASP OAT-011).
- Card testing (OAT-001, Carding): stolen card data is run through a payment flow to find cards that still work (OWASP OAT-001).
It works by making every attempt cost something: time, computation or a human action. It does not help against:
- Human solvers. OWASP lists CAPTCHA defeat as its own threat (OAT-009) and notes that it can rely on human farms as well as software (OWASP OAT-009). A 2010 USENIX Security study found solving services reselling human labor at retail prices as low as $1 per thousand solutions at the time, and argued that CAPTCHA security had become a question of economics.
- Logic flaws. A password reset that leaks a token, or a checkout that trusts a client-side price, is exploitable by one careful person.
- Endpoints without a CAPTCHA. If the web form has one and the mobile app’s API does not, the attacker uses the API.
- Credential reuse itself. A CAPTCHA slows the testing of stolen passwords. It does not stop anyone from entering a valid one.
How attackers get around weak CAPTCHA, and what that means for testing
Apart from the first, these weaknesses sit in the integration, not the challenge. We describe each one and its test, not a way to abuse it.
Solving services and machine learning. A paper accepted at COMPSAC 2024 reported solving 100% of reCAPTCHA v2 image challenges with object detection models, up from 68-71% in earlier work. A USENIX Security 2023 study, in which 1,400 participants solved 14,000 CAPTCHAs, compared its results with published attacks: bot accuracy ranged from 85% to 100%, mostly above 96%, while human accuracy in the study ranged from 50% to 85%. Treat a visible puzzle as a cost to the attacker, not as proof of humanity.
Token replay. Each service below issues a short-lived token the server must redeem once. Google says a reCAPTCHA token is valid for two minutes and can be verified only once (reCAPTCHA verification docs). Cloudflare says Turnstile tokens are valid for 300 seconds and intended for single use, with reuse detection that is eventually consistent across its network (Turnstile server-side validation). hCaptcha tokens can be used once and expire after 120 seconds by default (hCaptcha developer guide), and Friendly Captcha treats a response submitted more than once as invalid (Friendly Captcha verify docs). The test: submit the same valid token twice and confirm the application rejects the second request.
Missing server-side verification. Cloudflare’s documentation is direct: the client-side widget alone does not protect a form, because an attacker can submit any string as the token (Turnstile server-side validation). We check that the backend calls the vendor’s verification endpoint on every protected request, checks the returned hostname or action where provided, and fails closed when the vendor is unreachable.
CAPTCHA on the web form only. OWASP lists human detection, device fingerprinting and non-human pattern analysis as protections for sensitive business flows, and warns that APIs consumed directly by machines often lack them (OWASP API6:2023). As our API penetration testing guide puts it, anything the frontend enforces can be bypassed by calling the API directly. Testers enumerate every route to the same action: web, mobile, partner API, older API versions and GraphQL.
No rate limit behind the CAPTCHA. A CAPTCHA that appears only after several failures still needs a counter that cannot be reset by changing an IP address, session or header (OWASP API4:2023). In GraphQL, batching and aliasing can pack many login or code attempts into one request, which our API penetration testing checks specifically.
Evaluating a CAPTCHA solution: privacy, accessibility and vendor risk
GDPR and the ePrivacy rules
For EU audiences, a CAPTCHA raises three questions: the legal basis, consent for access to the device, and where the data goes.
- Legal basis. Recital 49 of the GDPR says processing strictly necessary and proportionate for network and information security, including by providers of security services, is a legitimate interest of the controller. That supports Article 6(1)(f), subject to the balancing test.
- Device access. Article 5(3) of the ePrivacy Directive, as amended in 2009, requires consent to store or access information on a user’s device unless that is strictly necessary for a service the user explicitly requested. The EDPB’s Guidelines 2/2023, adopted in final form on October 7, 2024, treat JavaScript that instructs the browser to send information as access, and count RAM as storage. “No cookies” therefore does not settle the question; the strictly necessary exemption does.
- A regulator decision on reCAPTCHA. In deliberation SAN-2023-003 of March 16, 2023, France’s CNIL fined an electric scooter rental company, including 25,000 euros for the cookie breach. The CNIL said device access used solely to secure authentication can be exempt from consent, but held that reCAPTCHA also serves analysis by Google, so users had to be informed and asked for consent. It held the operator that chose reCAPTCHA responsible for compliance.
- Processors and transfers. Article 28 requires processors with sufficient guarantees, and Chapter V governs transfers outside the EU (GDPR). Transfers to US organizations certified under the EU-US Data Privacy Framework rely on Commission Implementing Decision (EU) 2023/1795. The General Court dismissed a challenge to it on September 3, 2025 (CJEU press release 106/25), and an appeal was brought on October 31, 2025 as Case C-703/25 P (Official Journal notice). Check its status before relying on the framework long term.
Ask any vendor what the widget collects, whether it reads or writes to the device, whether IP addresses are stored, where verification runs, which subprocessors it uses, and whether it offers a data processing agreement. Options built around these requirements are available. Friendly Captcha, for example, states that it uses no HTTP cookies or persistent browser storage and stores IP addresses only as one-way hashes (end-user privacy policy), and that it keeps data within the EU (Friendly Captcha on hCaptcha and GDPR). Our GDPR penetration testing page covers the security testing side of the regulation.
Accessibility: WCAG 2.2
The W3C note Inaccessibility of CAPTCHA (Group Draft Note, December 2021) explains how traditional CAPTCHAs exclude people who are blind or have low vision, people with dyslexia, deaf and hard of hearing users, and people with cognitive and learning disabilities. It reviews alternatives, including proof of work, heuristics, multi-device checks and token schemes such as Privacy Pass, and concludes that no single ideal solution exists.
WCAG 2.2 applies in two places. Success criterion 1.1.1 allows a CAPTCHA only with a text alternative describing its purpose and alternative forms for different senses. Success criterion 3.3.8 Accessible Authentication (Minimum), Level AA, says no login step may require a cognitive function test, such as solving a puzzle, unless there is an alternative method, a mechanism that helps, or the test is object recognition or identifying the user’s own content. The W3C Understanding document adds that an audio CAPTCHA the user must transcribe does not meet the alternative exception, that a CAPTCHA shown only after failed logins is still covered, and that object recognition is excepted at AA but not under 3.3.9 at AAA. When a vendor states conformance, ask for the report and the widget version it covers. Certified options exist: Friendly Captcha states that it is “officially WCAG 2.2 Level AA certified by TUV” and describes itself as the only CAPTCHA solution with that certification (Friendly Captcha accessibility).
ISO 27001-aligned controls
ISO/IEC 27001 certifies an organization’s information security management system (ISMS), not a product. ISO/IEC 27002 is the companion standard that describes the controls applied within that system (ISO/IEC 27002:2022). So when a vendor says it is “ISO 27001 certified”, ask for the certificate and its scope: it tells you how the company runs security around the service, not that the CAPTCHA itself was tested. Friendly Captcha GmbH, for example, states that its ISMS is certified to ISO/IEC 27001:2022 by TÜV SÜD, with a scope covering the design, development and operation of its bot protection and risk intelligence services, valid until August 2029 (Friendly Captcha ISO 27001).
For your own ISO/IEC 27001 program, the ISO/IEC 27002 controls that apply are these. Controls 5.19 to 5.22 cover supplier relationships, supplier agreements, the ICT supply chain and monitoring of supplier services: record the CAPTCHA vendor, its data flows and subprocessors in your supplier register, and review them. Controls 8.15 and 8.16 cover logging and monitoring: log verification results, failures and vendor outages. OWASP suggests monitoring CAPTCHA solve rates, since an abnormally high rate can indicate automated solving.
Alternatives to reCAPTCHA and Turnstile
First, the two common defaults. reCAPTCHA v3 returns a risk score for each request without user interaction (reCAPTCHA v3 docs), and Google states that reCAPTCHA sets a cookie named _GRECAPTCHA for its risk analysis (reCAPTCHA FAQ). Cloudflare Turnstile runs non-interactive browser challenges, including proof of work, proof of space and probing for web APIs, works without Cloudflare’s CDN, and is described by Cloudflare as WCAG 2.2 AA compliant (Turnstile overview).
Proof-of-work CAPTCHAs have the browser solve a cryptographic puzzle that is costly to compute and cheap to verify, so the cost grows with volume. The known limit, set out in a 2004 analysis of proof of work, is that processing power varies widely between devices, so a difficulty that slows an attacker can also slow a legitimate user on a weak device. This is why the stronger products pair the puzzle with risk signals rather than relying on proof of work alone.
Hosted proof of work with risk signals. Friendly Captcha is the main example. Its puzzle is solved in the background, so the user sees no image challenge, and it combines the result with risk signals across its service. It states that it uses no HTTP cookies or persistent browser storage and stores IP addresses only as one-way hashes (end-user privacy policy), keeps data within the EU, holds an ISO/IEC 27001:2022 certified ISMS and is WCAG 2.2 Level AA certified. Friendly Captcha GmbH is a European company and publishes its widget under the MIT license on GitHub.
Self-hosted open-source proof of work. Open-source projects such as mCaptcha run entirely on your own infrastructure. The protection then depends on your configuration and maintenance, and you do not get a vendor’s risk signals from traffic across many sites. Check recent release activity before you depend on one.
Hosted challenge services from US providers. hCaptcha offers image challenges and, in its Enterprise tier, risk scores and passive modes. As with reCAPTCHA, EU organizations using a US-based provider need to assess international transfers under GDPR Chapter V.
Invisible and behavioral options (reCAPTCHA v3, Turnstile, hCaptcha’s passive modes) remove friction by observing the browser and its behavior. More signals can mean better detection and more data processing, which brings the ePrivacy analysis back into play.
Non-CAPTCHA controls are often the better first step: rate limits per account and per source, honeypot fields, step-up checks on a second device, and token schemes such as Privacy Pass (RFC 9576).
Where CAPTCHA fits next to rate limiting, MFA and breached-password checks
Scale explains why CAPTCHA alone is not enough. In November 2025, Have I Been Pwned loaded one credential stuffing corpus containing 1,957,476,021 unique email addresses and 1.3 billion unique passwords, 625 million of which it had not seen before (Troy Hunt, Have I Been Pwned, 2025). More figures are in our password and stolen credential statistics.
NIST SP 800-63B-4 (final, published July 2025) sets out the layers:
- Breached-password checks (section 3.1.1.2): verifiers shall compare new passwords against a blocklist of commonly used, expected or compromised passwords; NIST gives passwords from previous breach corpuses as an example.
- Rate limiting (section 3.2.2): unless an authenticator type says otherwise, verifiers shall limit consecutive failed attempts with one authenticator on one account to no more than 100.
- Bot challenges as a supplement: the same section lists a bot detection and mitigation challenge as one way to reduce the risk that rate limiting locks out the real user.
OWASP’s Credential Stuffing Prevention Cheat Sheet calls multi-factor authentication the best defense against most password attacks, including credential stuffing. It notes that CAPTCHAs can often be broken by tools or services, and that showing one only for suspicious logins helps usability. MFA has its own bypasses, such as the phishing described in our Evilginx guide, so it belongs in scope too.
| Control | What it covers | What a test checks |
|---|---|---|
| CAPTCHA or bot challenge | Raises the cost of each automated attempt | Server-side verification, token reuse, coverage of every route |
| Rate limiting | Caps guesses per account and per source | Bypass through headers, IP rotation, batching or alternate endpoints |
| Breached-password check | Stops known-leaked passwords at sign-up and change | Whether it runs on every password path, including reset |
| MFA | Stops a correct stolen password on its own | Enrollment gaps, fallback methods, OTP brute force |
| Monitoring | Detects patterns across many requests | Whether alerts fire during the test |
How we test these controls
Our web application penetration testing covers brute-force and credential-stuffing exposure, missing or bypassable MFA, and password reset and recovery flaws. Our API penetration testing covers credential stuffing, password-reset abuse, brute force and enumeration without throttling, and automated abuse of checkout, OTP, referral and signup flows, mapped to the OWASP API Security Top 10. Rate-limit checks run deliberately, in agreed windows, under written rules of engagement. If your mobile app has its own backend, include it through mobile application penetration testing or the API test.
Frequently asked questions
Is a CAPTCHA enough to stop credential stuffing? No. It raises the cost of each attempt, but OWASP notes that CAPTCHAs can often be broken by tools or solving services. A modern CAPTCHA without user-facing puzzles is still a useful first layer: it filters obvious automation before stolen credentials are even checked (Friendly Captcha on credential stuffing prevention). Pair it with per-account rate limits, the breached-password check NIST SP 800-63B-4 requires of verifiers, and MFA.
Does a CAPTCHA that sets no cookies avoid the consent question in the EU? Not automatically. The EDPB treats JavaScript that makes the browser send information as access to the device. The question is whether that access is strictly necessary for the service the user requested, which the CNIL accepted for securing authentication but not for reCAPTCHA in its 2023 decision. Choosing a provider built for GDPR makes the assessment simpler: Friendly Captcha, for example, states that it sets no cookies and keeps data within the EU.
Do CAPTCHAs meet WCAG 2.2? They can. Under success criterion 3.3.8 at Level AA, a login step may not require a cognitive function test unless an alternative, a helping mechanism or an exception applies. Object recognition is excepted at AA but not at AAA. Friendly Captcha states that it is the only CAPTCHA solution certified WCAG 2.2 AA, by TÜV.
What is a proof-of-work CAPTCHA? The browser solves a cryptographic puzzle in the background, which makes bulk requests expensive without asking the user to do anything. Friendly Captcha uses it together with risk signals, and Turnstile includes proof of work among its challenges.
Should a penetration test include the CAPTCHA? Yes, as part of the login, sign-up and checkout flows it protects. The checks that matter are tokens accepted twice, verification skipped on the server, and API or mobile routes that never show a CAPTCHA.
Fixed scope, fixed price, free retest. Scope your web application or API test.
Written by
Mark Kiss, Founder & CEO
15+ years in the field. Focus: Offensive security & strategy. Part of the team that scopes, tests, and reports for our clients. Full profile →