They built CyberDual.
AI attackers hide working exploits in real code that agentic reviewers often miss
CyberDual pits six attacking agents against three deployed code-review agents, judged by whether the exploit still runs after review, and finds open-weight GLM 5.3 is the hardest to stop.
Academic
Chenyuan Yang · Hwiwon Lee · Jongseong Kim · Yihan Yang · Dongjun Kim · Yinfang Chen · +1 more
University of Illinois Urbana-Champaign · Arizona State University
Research Digest··3 min read
The authors introduce CyberDual, an arena where AI attacker agents plant exploitable bugs in real maintenance commits to 14 open-source C/C++ libraries and AI reviewer agents try to find and repair them.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.