Exploit Generation

Every piece of software can contain bugs. Some of these bugs can be exploited: an attacker can use them to break into a system, steal data, or cause damage. The finished code that does exactly that is called an exploit – roughly “something that exploits.” Exploit generation is the process of creating this attack code. In the past, this was manual work for experienced security researchers. Today, AI systems – especially large language models, programs that can understand and generate text and code – can accelerate or partially automate this process. This fundamentally changes the threat landscape, because it drastically lowers the effort required for an attack.

Why exploit generation is putting pressure on security research

Often only hours pass between a vulnerability becoming known and the first real attack. In the security industry, this time window is called the “window of exposure.” The faster an exploit is created, the shorter this window – and the less time remains to update systems.

The real problem is the asymmetry: a defender must close every single gap in the system. An attacker only needs to find and exploit a single one. Automated exploit generation reinforces this imbalance, because it also allows less experienced attackers to attack complex vulnerabilities. Where extensive knowledge of assembly language, memory management, and operating systems was once required, today one can simply query an AI tool. The barrier to entry falls, and the attack surface grows.

How automated exploit generation works technically

The process typically runs in several stages. First, a system analyzes the source code or the compiled version – i.e., translated into machine code – of a piece of software and searches for patterns that indicate bugs, such as certain ways memory is managed. Then it attempts to deliberately trigger this bug.

A classic technique for this is fuzzing: the system sends massive amounts of randomly altered inputs to a program and observes whether it crashes. AI models make fuzzing more efficient because they learn from previous crashes and generate more targeted inputs instead of guessing blindly. In the final step, a finished exploit is built from the discovered bug – code that exploits the bug in such a way that the attacker gains a concrete advantage, for example elevated privileges on the system.

Researchers at the University of Illinois demonstrated in 2024 that a large language model was able to independently identify vulnerabilities in real, publicly known systems and write working exploits for them – without human assistance in the actual attack step. This is no longer a theoretical experiment, but evidence that the capability already exists.

Exploit generation in news, tools, and the security industry

In legitimate security research, exploit generation is actively used. Companies pay so-called penetration testers to attack their own systems in order to find vulnerabilities before criminals do. Tools like Metasploit, a widely used platform for security testing, contain collections of ready-made exploits for known vulnerabilities. AI-powered extensions of such tools are under active development.

In media coverage, the term mainly appears when AI systems achieve surprisingly good results in automatically attacking systems – whether in security competitions known as capture-the-flag contests or in research publications. The discussion around so-called bug bounty programs, in which companies pay cash rewards for discovered vulnerabilities, is also increasingly centered on the question of whether AI-generated exploits may be submitted.

A common misconception is that exploit generation is automatically illegal. That is not true. Its legality depends on against whom it is used and with what permission. The very same code is a tool of defense when used on behalf of a company – and a crime when used against someone else’s systems without authorization.