Capture the Flag

Capture the Flag is a sport-like competition centered on computer security. Organizers deliberately build flawed programs, websites, or servers. Whoever exploits one of these flaws and penetrates deep enough into the system finds a hidden string of characters there. This string is called a "flag" and looks something like this: flag. You enter it on the competition's website and earn points. The name comes from the outdoor game in which two teams try to capture the opposing side's flag.

Why companies send their best people to CTFs

Computer security is hard to learn from books. You have to actually experience how an attack unfolds in order to later be able to defend against it. That's exactly what these competitions are for: they provide an environment in which you're allowed to attack without harming anyone. Everything runs on the organizers' isolated servers.

For companies, CTFs are therefore a kind of talent market. Anyone who finishes near the top in a major international competition has essentially proven their skills. A diploma can hardly demonstrate that. Some security firms are made up almost entirely of people who got to know each other in this scene.

For some years now, there's been a second reason. CTF challenges have become a popular benchmark for AI systems. They have a clearly correct solution—namely the flag—and can therefore be evaluated automatically. At the same time, they require several reasoning steps in sequence. A model has to understand a program, suspect a vulnerability, build an attack, and try it out.

From the challenge to the flag

Most competitions divide their challenges into categories. "Web" is about flaws in websites, such as a login that can be bypassed with a crafty input. In "Reverse Engineering," you're given a finished program without the corresponding source code and have to figure out which password it accepts. In "Crypto," encryption has been implemented incorrectly and can therefore be cracked. Then there's "Forensics": you search through files or captured network traffic for traces.

There are two basic formats. In the Jeopardy format, all teams work independently on the same list of challenges, with each challenge worth a different number of points. In the Attack-Defense format, each team gets its own server with the same vulnerabilities. You have to patch your own holes while simultaneously exploiting those of the others. This is considerably more frenetic and often runs for 24 or 48 hours straight.

A clear distinction from real hacking is important. A CTF flag has no value outside the competition, and the systems belong to the organizer. Anyone who applies the same techniques to someone else's servers is committing a crime. This distinction isn't a minor detail—it's the fundamental rule of the entire scene.

CTFs as a yardstick for AI agents

CTF challenges now appear regularly in reports about AI security. Labs such as OpenAI, Anthropic, and Google DeepMind have their models solve such challenges before release. The results then show up in the model's safety reports. The reasoning behind this: if a system can crack difficult CTF challenges on its own, it might also be able to automate real attacks.

The numbers are moving fast. Just a few years ago, models still failed at beginner-level challenges. Today, the strongest systems solve a significant portion of medium-difficulty challenges on their own, provided they're allowed to use tools like a command line. On the hardest challenges—the ones that keep even skilled humans busy for hours—they still often fail.

For pupils and students, CTFs are also a very low-barrier entry point. Platforms like picoCTF are explicitly aimed at beginners and run entirely in the browser. You don't need your own server or expensive software. In Germany, there are also competitions at universities as well as the "Cyber Security Challenge" for young people.