Model Denial of Service

What is Model Denial of Service?

Model Denial of Service (DoS) occurs when an attacker interacts with an LLM in a way that consumes an excessive amount of resources (CPU, GPU, memory, or context window), leading to service degradation or unavailability for other users.

Common attack vectors include:

• Context Window Overflow: Sending extremely long prompts that fill up the model's context window.
• Recursive Expansion: Asking the model to perform tasks that generate exponentially larger outputs.
• Variable-Length Input Flooding: Sending a flood of requests with varying lengths to disrupt batching optimizations.