Exploit

Every larger program contains bugs. Most of the time you don’t notice anything, sometimes an app crashes. But some of these bugs can be abused: whoever sends the right input can make the program do something completely different from what was intended. This specially crafted input is exactly what’s called an exploit. The English word means roughly “to take advantage of.” The flaw itself is called a vulnerability or security hole, the exploit is the matching key to it.

From programming error to real damage

A vulnerability alone doesn’t cause any damage. It’s only a theoretical possibility. Only the exploit turns it into a practical attack. That’s why experts strictly distinguish between the two. A vulnerability for which no one has a working exploit is considered far less urgent.

So-called zero-day exploits are especially feared. This means: the vendor doesn’t yet know about the flaw and had zero days to fix it. So no update exists. Such exploits are traded on black markets for six-figure sums, sometimes bought up by intelligence agencies as well.

The damage can be enormous. In 2017, the ransomware WannaCry spread via an exploit for Windows and paralyzed hospitals and corporations worldwide. An update already existed at the time, but many systems hadn’t been updated. This shows: the race between attackers and defenders is often decided by how quickly updates are installed.

What happens when a flaw is exploited

Many exploits take advantage of the fact that a program doesn’t cleanly separate data and commands. A classic example is the buffer overflow: a program reserves space for 100 characters, the attacker sends 500. The excess characters end up in memory areas that are actually meant to hold instructions. The attacker writes their own instructions there, and the computer dutifully executes them.

A comparison makes the principle tangible. Imagine a form where, instead of a name, someone writes an instruction to the clerk in the name field. If the clerk doesn’t treat the text as a name but instead follows it as an order, the form is insecure. This exact logical flaw underlies a great many exploits.

The goal is almost always one of two things. Either the attacker gains more privileges than they’re entitled to, such as administrator rights. Or they get foreign code to execute on the targeted machine. Modern operating systems defend against this with protective mechanisms, for example by arranging memory differently on every startup. Successful attacks therefore often chain together several exploits in sequence.

Exploits in AI systems and in the news

Exploits regularly show up in the news when Apple, Google, or Microsoft release an emergency update. The note “already being actively exploited” then means that a working exploit is circulating in the wild. For users this simply means: update immediately. A software company’s stock price can also suffer when a severe vulnerability becomes public.

Not every exploit is written by criminals. Security researchers build them deliberately to prove that a vulnerability is real. Such demonstrations are called proof of concept. Companies pay rewards for such reports, so-called bug bounties. At competitions like Pwn2Own, there are large cash prizes for teams that break into current software.

With AI systems, the picture shifts. Language models can’t be attacked via memory errors, but via cleverly worded text. This is referred to as prompt injection: a hidden instruction in a website or email causes the AI assistant to ignore its rules. The underlying principle is the same as with a classic exploit. Once again, a system fails to properly separate data and commands.