Function Hijacking Attack Threatens Agentic AI Models

A new paper on arXiv (2604.20994) introduces a function hijacking attack (FHA) that targets function calling in large language models (LLMs) used in agentic AI systems. Unlike existing attacks that focus on semantic preference manipulation, FHA forces the invocation of a specific attacker-chosen function by manipulating the tool selection process. The research highlights vulnerabilities in agentic models' function calling interfaces, which can lead to data tampering, theft, endless loops, or harmful content generation. The attack expands on known injection and jailbreaking methods, demonstrating a novel threat vector for AI-powered systems that rely on external function invocation.