Supervisors

Suitable for

Abstract

Agentic systems, powered by large language models (LLMs), are increasingly deployed to perform actions such as navigating the internet, booking hotels, scheduling flights, and managing calendars. These agents often rely on datasets derived from API interactions to make decisions. However, this reliance makes them vulnerable to adversarial API injection attacks, where malicious actors manipulate API call responses to bias agent behavior. This project explores how adversarial API calls can cause LLM-based agents to disproportionately favor certain APIs, raising significant fairness concerns—particularly as API developers are often compensated based on usage metrics. The project will design attack strategies, evaluate their impacts on fairness and functionality, and propose robust mitigation techniques. This project is designed to lead to a high-quality publication, and we are looking for a highly motivated student to contribute to this groundbreaking area of research.

 

We probably will be able to collaborate with the UK’s AI Safety Institute. Moreover, we will work closely with Guohao Li (co founder of Eigent ai and Camel AI -- London based startup).

 

[A] CRAB: Cross-environment Agent Benchmark for Multimodal Language Model Agents. (https://arxiv.org/pdf/2407.01511)