AI agent security
Who decides what an agent may do?
Study authorisation across roles, resources and actions, including least privilege, tool access and human confirmation, so an agent’s capabilities and responsibilities can be examined.
代理人紅隊攻防實境實驗室
Department of Computer Science and Engineering, Yuan Ze University
Understand the attack.
Evaluate the defence.
AI AGENTS · RED TEAMING · ZERO TRUSTMIRAGE Lab is based in the Department of Computer Science and Engineering at Yuan Ze University. Assistant Professor Wilbur Wei (Te-En Wei) is its Director. Its research focuses on AI agent security, red teaming and zero trust, examining how attack scenarios, authorisation boundaries and behavioural evidence inform the design and evaluation of defences. Future directions include drone and robot communication security and PUF hardware security applications.
DirectorWilbur Wei / Te-En WeiAssistant Professor, Computer Science and Engineering, Yuan Ze University
Faculty backgroundWho decides what an agent may do?
Study authorisation across roles, resources and actions, including least privilege, tool access and human confirmation, so an agent’s capabilities and responsibilities can be examined.
What evidence does an automated attack leave?
Examine AI-assisted security testing, encrypted lateral movement and host behaviour to understand how attacks can be observed, detected and assessed.
How can a system reassess access and account risk?
Connect continuous trust inference, Active Directory account risk and APT intelligence to access and defence decisions grounded in multiple sources of evidence.
基於 LLM 驅動的紅隊代理攻防演練與場域模擬實戰平台
View research projectBuilding on existing AI security and zero-trust research, I plan to deepen work in communication security and hardware security. The directions below are prospective research plans, open to discussion on problem definition, experimental design and industry collaboration.
How can devices, controllers and services establish trust in commands and data?
Planned work examines authentication, command authorisation and communication protection between drones, robots and control systems. Threat scenarios such as impersonation, replay and unauthorised control will inform evaluation of defences and their effects on latency, availability and operational safety.
How can hardware characteristics support device identity and key protection?
Explore Physical Unclonable Functions (PUFs) for device identification, authentication and key generation. Planned evaluation considers reliability under environmental variation, resource constraints and resistance to attacks, as well as connections to device and communication security.
These are prospective directions. Published papers and awards are listed separately.
Related work by the Director and co-authors includes AHAF, HMELF, MARS, STIE-ZTA, AD risk and APT intelligence analysis. The research page provides author lists, publication venues and the scope of each study.
For academic or industry research on AI agent security and zero trust, or planned work on drone / robot communication security and PUF applications, share the research question, relevant background and your collaboration idea.
Share your event or team context, the problem to address and the expected timeline.
Prepare an inquiry, then copy it or choose an email option to send it.