OWASP Top 10

Every website and every app that exchanges data over the internet can contain flaws that attackers can exploit. Such flaws are called security vulnerabilities. The OWASP Top 10 is a list that summarizes the ten most important types of such vulnerabilities. It is compiled by OWASP, a nonprofit organization in which software security professionals gather their knowledge and make it freely available. The list is not published anew every year, but roughly every three to four years, most recently in 2021 and in a revised version in 2025. It is not a law or a technical standard, but a recommendation that has established itself as a de facto standard in practice.

Why a list of ten items carries so much weight

Security is a topic with no natural end point. You can always test, secure, and monitor more. That is precisely the problem: those who don’t know where to start often don’t start at all. The OWASP Top 10 solves this by providing a clear order of priority. Anyone who works through the ten points has eliminated the errors that most frequently cause real-world damage.

There is also an economic aspect to this. Many companies require their software suppliers to take the OWASP Top 10 into account. As a result, the list constantly appears in tenders, contracts, and audit reports. Certifications and security audits also frequently rely on it. A development team that ignores the list has a harder time winning customers.

A common misconception: the Top 10 is not a complete checklist. It covers the most common risks, not all of them. Anyone who only checks off these ten points and then stops has a solid foundation, but not a secure application. OWASP itself explicitly emphasizes this and also publishes much more extensive testing standards alongside the list.

How the ten rankings are determined

The list is not negotiated in a round of experts, but is largely calculated from data. OWASP collects analyses from security firms that test applications for vulnerabilities on behalf of their clients. This produces data sets covering hundreds of thousands of applications. For each type of flaw, it is measured in how many applications it occurred and how severe the consequences typically were. The ranking emerges from this combination.

However, two spots are deliberately not assigned on a purely computational basis. They come from a survey of professionals. The reason is simple: automated tests only find what they are looking for. New types of attacks only appear in the data with a delay. The survey is meant to prevent the list from lagging behind reality.

The entries themselves are usually not individual bugs, but categories. One example is “Broken Access Control,” meaning flawed access control. This refers to a situation in which a user can view or modify data they are not authorized to access — for instance, when changing an invoice number in the address bar and suddenly seeing the invoice of a different customer. This category has topped the list for years.

From testing software to AI applications

In developers' everyday work, the list is often built directly into tools. Programs that automatically scan source code for errors assign their findings directly to an OWASP category. Reports from commissioned attack simulations, so-called penetration tests, are also usually structured according to these categories. In job postings for security positions, the list is a standard requirement.

In the news, the term usually comes up after major data breaches. When customer data leaks, the cause can almost always be traced back to one of the ten categories. Security researchers often frame this as an accusation: the flaw has been known for years and is, after all, listed in the OWASP Top 10.

A dedicated variant now exists for the AI field: the OWASP Top 10 for Large Language Models. It addresses risks that did not exist in this form on classic websites. At the very top is “Prompt Injection”: an attacker hides instructions within a text or on a website that the AI system then follows. Poisoned training data and the unchecked reuse of AI-generated responses are also listed there.