New benchmark measures whether coding AI refuses to write malware

Original: Code as a Weapon: A Consensus-Labeled Prompt Bank for Measuring Coding-Model Compliance with Malicious-Code Requests

Source: arxiv.org ↗