AI Agent Security: Attacks, Failures, and Open Problems

DETAILS

This is a hybrid event held in Christchurch, with an online option available.

This event will be delivered in Mandarin.

ABSTRACT

In recent years AI agents have significantly boosted productivity, with software engineering as the most prominent example. Compared with how fast applications have grown, the security of agent systems themselves remains under-explored.

This session starts by introducing an analytical framework for understanding and analysing agents. It characterises and classifies different types of agents along dimensions such as system composition and interaction patterns, giving us a common vocabulary before we talk about what goes wrong.

With that framework in hand, we turn to real cases and recent research to examine the attacks, vulnerabilities and failure modes that agent systems face. Some of these are practical risks already seen in deployed applications; others are open challenges that current research has not adequately addressed.

The aim is to help you see AI agent security from a more systematic perspective, and to leave with a sense of which directions are worth exploring further.

What we will cover:

• Framework: an analytical framework for describing agents by system composition and interaction patterns
• Taxonomy: how different types of agents can be characterised and classified within that framework
• Attacks: real-world attacks, vulnerabilities and failure modes in agent systems, illustrated with cases
• Research: what the latest work has found and what it has not yet solved
• Directions: open challenges and research directions worth pursuing

Whether you are building agents, integrating them into a product, or simply want to understand what could go wrong before it does, this session is for you.

SPEAKER BIO

Ke He is at the University of Canterbury, where his work focuses on the security of AI agent systems. This talk draws on that research, combining real-world cases with recent results to give a systematic view of how agents fail and how they are attacked.

AGENDA

18:00 – 18:30 Registration, light refreshments and networking
18:30 – 19:30 Presentation, with interactive Q&A throughout
19:30 – 20:00 Open Q&A and networking

LOCATION

Rarakau: Riccarton Centre
199 Clarence Street, Riccarton, Christchurch 8053

ADMISSION FEE

Free for CITA members
$5 donation for non-members
Account Number: 06-0513-0847960-00
Account Name: CHINESE IT ASSOCIATION NZ

NOTES

Light refreshments and drinks will be provided.

An online joining link will be shared with registered attendees before the event. The online session will include slides and audio only.