UNIST will lead a five-and-a-half-year research project to develop security technologies for agentic AI, with a focus on preventing malicious inputs from manipulating how AI agents make decisions and carry out tasks.

The project has been selected for the 2026 AI Star Fellowship , supported by the Ministry of Science and ICT (MSIT) and the Institute of Information & Communications Technology Planning & Evaluation (IITP). A consortium, led by UNIST will receive KRW 11 billion in government funding through December 2031, working with Sungkyunkwan University and three AI and cybersecurity companies.

Professor Hyungon Moon, Head of the Department of Computer Science and Engineering, will lead the project. The UNIST team also includes Professors Seongil Wi, Minkyung Park, Saerom Park, and Hyungho Na, alongside researchers from Sungkyunkwan University and industry partners SecuLayer, S2W, and RaonData.

Agentic AI systems can use large language models (LLMs) to read documents, search the web, access external tools, and carry out tasks. These capabilities also introduce security risks: malicious inputs can manipulate an agent's decisions and potentially influence the actions it takes.

Current defenses often protect individual components separately, for example by detecting malicious inputs or isolating execution environments. The research team will instead coordinate safeguards across the model, interface, and execution environment, creating multiple lines of defense so that an attack that bypasses one layer can still be contained at another.

The consortium will pursue three projects addressing security challenges across these layers and test the resulting technologies in services operated by its industry partners. Professor Moon's group brings expertise spanning systems security and AI, including trusted execution environments (TEEs), formal verification, knowledge distillation, and methods for evaluating the performance and limitations of deepfake defenses. Its systems security research has been presented at major international conferences including USENIX Security and the IEEE Symposium on Security and Privacy.

The project will also train early-career researchers working at the intersection of AI and cybersecurity, while seeking to expand the use of AI agents in settings where security concerns have limited their adoption.

“Securing agentic AI requires more than protecting the model itself; the pathways through which inputs enter the system and the environments in which actions are executed must also be protected,” said Professor Moo, Chair of the Department of Computer Science and Engineering. “Through this project, we aim to develop core security technologies spanning all three layers and train researchers who can help advance the safe use of agentic AI.”