Soongsil University Undergraduate Researcher's Paper Accepted to International Natural Language Processing Conference 'EMNLP 2026'

[Financial News]  Soongsil University announced on the 3rd that a paper co-authored by undergraduate researcher Kim Do-yoon as first author had been accepted to the main conference of The 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026).
Natural language processing is a field of artificial intelligence (AI) technology that enables computers to understand and process human language.
Undergraduate researcher Kim Do-yoon was the first author, while Chan-Woo Kim, an undergraduate researcher in the same laboratory, was a co-author. The study examined security vulnerabilities that may arise in self-evolving AI agents, which create and improve new functions on their own.
Self-evolving AI agents improve their performance by creating new functions based on past task experiences, storing them, and reusing them in later tasks. The researchers analyzed the possibility that erroneous or malicious behavior could be accepted as a normal experience during this process and stored as a function.
In other words, an erroneous function that has been stored once could be repeatedly used in subsequent tasks.
The researchers named this type of attack “EvoSkill Injection” and proposed an evaluation method called “SARGE” to assess related vulnerabilities.
This is a so-called “Red Teaming” approach. It involves assuming that an actual attacker is targeting the system and creating various attack scenarios to identify security weaknesses.
The researchers determined that simply examining what users input and what answers the AI produces is insufficient when assessing the safety of AI agents. They explained that the entire process must also be examined, including how the AI creates, verifies, and stores new functions before retrieving and using them again later.
Kim Do-yoon said, “It is meaningful to be able to present the results of our research on the safety and security of AI agents at EMNLP 2026. I would like to continue researching ways to build reliable and safe AI agents.”

[email protected] Bomi Lee Reporter