Security Threats in Modern LLM Applications

Explore critical security vulnerabilities and defense strategies for Large Language Model applications in this 20-minute conference talk from Conf42 LLMs 2025. Begin with an introduction to the unique security challenges posed by LLMs and understand how these AI systems differ from traditional applications in terms of attack vectors. Examine prompt injection attacks and learn how malicious inputs can manipulate model behavior, then investigate insecure output handling practices that can lead to system compromises. Discover the risks of training data poisoning and how contaminated datasets can affect model integrity, followed by an analysis of model theft and extraction techniques used by attackers. Understand excessive agency risks where LLMs are given too much autonomy, and explore how sensitive information disclosure can occur through model responses. Analyze supply chain vulnerabilities in LLM development and deployment pipelines, then examine the dangers of over-reliance on AI decision-making systems. Learn about denial of service attacks specific to LLM infrastructure and conclude with comprehensive best practices for securing LLM applications throughout their lifecycle.