Testing Russian LLMs Against OWASP LLM10: Unbounded Consumption Attacks

The article discusses the OWASP LLM10: Unbounded Consumption category, which focuses on attacks that force language models to perform disproportionately large amounts of computation, leading to resource exhaustion and denial of service. The researchers prepared malicious prompts representing various attack classes, including recursive context expansion (Quine-like recursion), fractal nesting, computational explosion, semantic combinatorial explosion, multi-agent simulation, and glitch token abuse. These prompts can cause models to generate excessively long responses, simulate thousands of entities, or recursively analyze their own reasoning, all while appearing as legitimate user requests. The study tested two Russian flagship LLMs, YandexGPT Lite 5 and GigaChat Lite, using the AI Red Team web service, which supports over 2000 malicious prompts, 200 of which target OWASP LLM10. The evaluation involved LLM-as-judge with manual validation. The article emphasizes that traditional security measures like keyword filtering are ineffective, and highlights the need for more robust architectural defenses.