✕

Security boundary

Denial of Service (DoS)

A Denial of Service (DoS) attack in the context of generative AI targets the system’s availability. The attacker aims to overload the model or its infrastructure—such as servers, APIs, or network resources—so that legitimate users cannot access it. This can be done by sending an excessive volume of requests, crafting computationally expensive queries that consume disproportionate amounts of resources, or exploiting inefficiencies in the model’s response generation.

Unlike content-focused attacks, DoS attacks do not necessarily try to extract secrets or produce harmful messages. Instead, their objective is to disrupt normal operations, degrade system performance, or shut down the service entirely. The impact is both operational (downtime, user dissatisfaction) and financial (lost business, remedial costs).

Example:

Imagine a public AI-powered customer service tool that handles thousands of user queries daily. An attacker writes a script that repeatedly sends extremely long, complex prompts that cause the model to run intense computations. Over time, the system’s servers become overloaded and slow down, eventually timing out for legitimate users. Frustrated customers cannot get their queries answered, causing reputational damage and forcing the company to deploy additional resources to restore normal service levels.