2608.15092

Total: 1

#1 WeSCE: A Benchmark for Measuring Security Drift in LLM-Driven Code Editing [PDF] [Copy] [Kimi] [REL]

Authors: Zhiyu Zhang, Tingyue Wen, Senke Sun, Dengxiang Liang, Enhao Huang

In this work, we introduce WeSCE, a benchmark for quantifying security drift in code editing under weak-security constraints, where tasks specify only functional objectives without explicit security requirements. WeSCE consists of 400 executable programs derived from real-world code, covering feature addition, feature removal, bug fixing, and refactoring. To quantify security drift, we propose a continuous risk representation that aggregates heterogeneous vulnerability signals through a unified formulation, and define drift measures capturing changes in overall risk, worst-case severity, and vulnerability distribution under code transformations, providing a multi-scale view of security spanning average-case behavior to worst-case emphasis.

Subjects: Cryptography and Security , Artificial Intelligence , Software Engineering

Publish: 2026-08-15 07:24:13 UTC