Таксономия Corsar: классификация методов обхода ограничений в языковых моделях по механизму воздействия
Year: 2026
Type: article
Abstract: Attacks aimed at bypassing the built-in security mechanisms of large language models to obtain responses to prohibited queries, also known as jailbreak attacks, are a critical and rapidly evolving threat. With the digital industry's transition to the use of complex AI agents, the danger of such attacks has increased significantly. Simple generation of prohibited text has given way to the risk of compromising information systems. Improving security mechanisms requires structuring attack methods. Existing attack taxonomies treat queries as indivisible units, classifying them solely by external characteristics. This leads to overlapping categories and fails to explain the mechanisms by which security is bypassed. The goal of this paper is to develop a new classification of attacks on large language models. We propose the CORSAR taxonomy, based on decomposing attacks into atomic techniques based on the mechanism by which they affect the model. If an attack combines several techniques, it is decomposed into parts, each of which is considered an independent unit of analysis. The groupings are based on the nature of the technique's operation. Using the attacker's techniques as a basis for classification minimizes the overlap between categories and makes them close to orthogonal. The taxonomy identifies six independent categories, divided into 20 subcategories, each containing a set of specific techniques. The proposed classification will allow us to create a threat profile for a large language model, which will make it possible to develop more flexible protection mechanisms.
Language: Russian
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7225364388