What is LLMDoS?
LLMDoS, told as a story about a few small tickets that keep the stove burning until the whole kitchen stalls.
One small ticket can tie up the whole kitchen for hours.Just one line of text — but the work it demands is a mountain.
You can't tell a big legitimate order from a costly one just by reading it.What matters is how much work it demands — and that only shows up once cooking has already begun.
LLMDoS is an attack that floods the kitchen with costly orders until it stalls.Not a truckload blocking the door — just a few tickets that burn out the stove.
DDoS stacks trucks at the door. LLMDoS keeps lighting the stove.Ordinary DDoS pushes with sheer traffic volume — LLMDoS exhausts resources through the compute cost a single request demands.
A fence that only counts tickets isn't enough to stop it.The rate limit has to measure cost, not just count — and this gap is wider precisely because the chef trusts every order as written, never suspecting on its own that one is too costly.
LLMDoS = an attack that keeps the kitchen's stove burning with just a few tickets until it stalls. Instead of stacking trucks, it exhausts the kitchen by demanding costly computation over and over.
LLMDoS (LLM Denial of Service) is an attack that exhausts a model's compute, memory, or time budget by sending long, repetitive, or recursively expanding requests. Unlike traditional DDoS, which pushes with sheer network traffic volume, LLMDoS can degrade or stall a service with only a handful of requests by triggering expensive token generation or long-context processing. Token/time limits, per-request cost estimation, and fair queuing are the main defenses.
When grown-ups say it
- LLMDoS
- The order that keeps the stove burning. Just a few tickets can stall the whole kitchen.
- Compute-bomb ticket
- A short note that demands a mountain of work. A recursively expanding request is the classic shape of one.
- vs. DDoS
- Stacking trucks at the door. The general attack that floods network traffic. → the security world's DDoS
- Rate limiting
- Lining up incoming tickets. It has to gate on cost, not count, to stop LLMDoS. → the theme park's rate limiting
- Availability
- Guests getting served in time. It's the first thing to break when the stove stalls. → the theme park's reliability
- Token limit
- The max number of beans usable at once. It caps the size of the ticket itself ahead of time. → the parrot world's beans
- Cost-based throttling
- A fence that measures work, not just count. Even the same number of tickets can carry wildly different costs.
- Capacity planning
- Figuring out ahead of time how many stoves can run at once. You have to size for both normal and peak load ahead of time. → the theme park's capacity planning