ML Security PapersBreaking MCP with Function Hijacking Attacks: Novel Threats for Function Calling and Agentic Models
Yannis Belkhiter 1,2, Giulio Zizzo 1, Sergio Maffeis 3, Seshu Tirupathi 1, John D. Kelleher 2
Published on arXiv
2604.20994
Input Manipulation Attack
OWASP ML Top 10 — ML01
Insecure Plugin Design
OWASP LLM Top 10 — LLM07
Excessive Agency
OWASP LLM Top 10 — LLM08
Key Finding
Achieves 70% to 100% attack success rate across 5 function-calling models including instructed and reasoning variants on BFCL dataset
Function Hijacking Attack (FHA)
Novel technique introduced
The growth of agentic AI has drawn significant attention to function calling Large Language Models (LLMs), which are designed to extend the capabilities of AI-powered system by invoking external functions. Injection and jailbreaking attacks have been extensively explored to showcase the vulnerabilities of LLMs to user prompt manipulation. The expanded capabilities of agentic models introduce further vulnerabilities via their function calling interface. Recent work in LLM security showed that function calling can be abused, leading to data tampering and theft, causing disruptive behavior such as endless loops, or causing LLMs to produce harmful content in the style of jailbreaking attacks. This paper introduces a novel function hijacking attack (FHA) that manipulates the tool selection process of agentic models to force the invocation of a specific, attacker-chosen function. While existing attacks focus on semantic preference of the model for function-calling tasks, we show that FHA is largely agnostic to the context semantics and robust to the function sets, making it applicable across diverse domains. We further demonstrate that FHA can be trained to produce universal adversarial functions, enabling a single attacked function to hijack tool selection across multiple queries and payload configurations. We conducted experiments on 5 different models, including instructed and reasoning variants, reaching 70% to 100% ASR over the established BFCL dataset. Our findings further demonstrate the need for strong guardrails and security modules for agentic systems.
Key Contributions
- Novel function hijacking attack (FHA) that forces LLM agents to invoke attacker-chosen functions by inserting adversarial tokens in function descriptions
- Demonstrates universality across queries and payloads, achieving 70-100% attack success rate on 5 different function-calling models over BFCL dataset
- Shows the attack is context-agnostic and robust across diverse domains, unlike existing semantic-preference-based attacks
🛡️ Threat Analysis
The attack uses gradient-based adversarial perturbations (adapting GCG) to manipulate model outputs at inference time — specifically inserting adversarial tokens in function descriptions to force specific function calls.