SK Shielders "AI Can Be Hacked Too"... Publication of LLM Vulnerability Diagnosis Guide

Supporting the Construction of Safe AI Systems

SK Shieldus announced on the 8th that it has published the ‘LLM (Large Language Model) Application Vulnerability Diagnosis Guide’ to proactively respond to artificial intelligence (AI) security threats.

LLM-based applications utilize large-scale language models specialized in natural language processing and generation, and are used in various industries such as finance, manufacturing, and healthcare. However, these applications are vulnerable to security threats different from those of existing IT systems due to the unique characteristics of data and user input processing methods, requiring thorough preparation.

SK Shieldus identified an increase in AI-based hacking as one of the major security threats in 2025, particularly expecting intensified hacking targeting small language models (sLLM) and data manipulation and leakage attacks exploiting structural vulnerabilities of LLMs. According to the report, representative security threats include ‘prompt injection,’ ‘API parameter tampering,’ and ‘RAG (Retrieval-Augmented Generation) data contamination.’

The guide emphasized the need to separate users and system commands (prompts) and strengthen data verification procedures to prevent these security threats. It also recommended using sandboxes to prevent malicious code execution depending on whether LLMs execute code, and establishing group-based permission management systems to block unauthorized data access. In particular, it added that a multi-layered security system should be introduced to prevent data contamination and privilege escalation attacks.

Kim Byung-moo, Vice President and Head of Cybersecurity Division at SK Shieldus, said, “AI technology provides convenience, but if security vulnerabilities caused by technical instability are exploited, serious hacking incidents can occur,” adding, “This guide will provide practical help in preventing AI security issues that companies and institutions may face in advance, while building trustworthy AI systems.”