V3AST-Bench contains 212 risk tasks spanning 18 categories and nine risk types, each paired with a capability twin task that preserves the workflow under benign conditions.
V3AST-Bench measures coding agent safety through verified execution outcomes.
Authors pair 212 adversarial coding tasks with benign capability twins to separate avoiding harm from inability to do the work.
Independent
Weitao Feng · Hengheng Xiong · Zihao Zeng · Jun Chen · Jiyang Guan · Xiaofeng Wang · +3 more
Research Digest··3 min read
The authors introduce V3AST-Bench, a benchmark for evaluating the safety of LLM-based coding agents using executable checks over environment states, records, and artifact behavior.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.