OWASP LLM10: Unbounded Consumption — Testing Modern Language Models on Resource-Intensive Prompts
Яндекс
Сбер
The article explores the OWASP LLM10 category (Unbounded Consumption) in the context of LLM security, describing attack classes that cause disproportionate resource consumption. The authors tested Russian flagship models YandexGPT Lite 5 and GigaChat Lite using the AI Red Team system, which includes over 2000 malicious prompts.
Large language models are increasingly used in corporate systems, but with their popularity, the threat profile changes: attackers can formulate requests that force models to perform resource-intensive tasks, consuming CPU time, memory, and thousands of tokens. The OWASP Top 10 for LLM Applications includes the LLM10: Unbounded Consumption category, which adapts the Denial of Service (DoS) idea to generative AI. Unlike classic attacks like Prompt Injection or Data Leakage, these attacks target computational resources. The article describes several attack classes: recursive context expansion (Quine-like Recursion), fractal nesting, computational explosion, semantic combinatorial explosion, multi-agent simulation, and glitch token abuse. The authors conducted experiments using the AI Red Team system, which supports over 2000 malicious prompts, 200 of which target LLM10. They tested two Russian flagship models: YandexGPT Lite 5 and GigaChat Lite. The study was supported by the Innovation Promotion Foundation (fasie.ru), and the examples are provided for research and educational purposes.
- Abbreviations
- LLM = Large Language Model — большая языковая модель
- OWASP = Open Web Application Security Project — Открытый проект безопасности веб-приложений
- API = Application Programming Interface — программный интерфейс приложения
- GPU = Graphics Processing Unit — графический процессор
- RAG = Retrieval-Augmented Generation — генерация с дополнением извлечением
- DoS = Denial of Service — отказ в обслуживании
- Unicode = Unicode — Юникод
Source: Habr — хаб ИИ —
original