Exploit

An exploit is attack code or a technique that actually takes advantage of a security hole already present in a computer system to break in. Think of a house: knowing a lock has a flaw is one thing, but building a tool that actually uses that flaw to open the door is something else entirely. The flaw itself is the vulnerability, while the key or tool that actually pries the door open using that flaw is the exploit.

This term has become important in AI safety discussions because whether a model can go beyond merely finding vulnerabilities and actually generate working attack code that exploits them has become a key criterion for assessing risk. If an AI can go all the way from discovering a vulnerability to writing intrusion code without human help, it means that AI could be far more dangerous as an offensive tool than as a defensive one.

In articles, it appears in phrasing like "it must be able to identify zero-day vulnerabilities and develop a functioning exploit without human intervention." A common misconception is treating "exploit" as synonymous with "finding a vulnerability," but it actually refers to the executable means of attack that uses that vulnerability to actually break into a system or trigger a desired behavior. Discovering a vulnerability and developing an exploit are sequential but distinct stages.