What question did this study set out to answer?

Investigate the vulnerabilities of large language models to attacks, particularly prompt injection.

February 26, 2026Automatic Control and Computer Sciences0 citations

From Exploitation to Protection: Analysis of Attacks on Large Language Models

SBS. V. BezzateevSaint Petersburg State University of Aerospace and Instrumentation IVI. S. VelichkoSaint Petersburg State University of Aerospace and Instrumentation

Key Points

Investigate the vulnerabilities of large language models to attacks, particularly prompt injection.
Analyzed various attack vectors targeting large language models.
Focused on the mechanism of prompt injection attacks.
Assessed the impact of these attacks on model performance and data security.
Identified prompt injection as a significant threat to model integrity.
Demonstrated that certain attacks can manipulate model behavior and leak confidential information.
Established that models can be coerced into following harmful instructions.

Abstract

Modern large language models possess impressive capabilities but remain vulnerable to various attacks capable of manipulating their responses, causing confidential data leaks, or bypassing restrictions. The main focus is on analyzing “prompt injection” attacks, which allow circumventing model limitations, extracting hidden data, or forcing the model to follow malicious instructions.

Demander à l'IA

Bookmark

View Full Paper