New Function Hijacking Attacks Threaten Agentic AI Models

Researchers have identified a new security threat called Function Hijacking Attacks (FHA) that targets agentic AI models utilizing function calling capabilities. These attacks manipulate the model's tool selection process to force the invocation of an attacker-chosen function, bypassing semantic understanding and remaining effective across different domains and function sets. The FHA has demonstrated a significant success rate on various LLMs, including instructed and reasoning models, and shows transferability across different model sizes and families, highlighting the urgent need for robust security measures in agentic AI systems. AI

IMPACT Highlights critical security vulnerabilities in agentic AI, necessitating new guardrails and security protocols for deployed systems.

RANK_REASON Academic paper detailing a new type of security attack on AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• agentic AI
• function calling
• Function Hijacking Attacks
• Hugging Face
• large-language models
• Yannis Belkhiter

• safety
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →