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

arXiv·medium signal

Argues that coding models returning working malicious code (keyloggers, ransomware stubs, exploits) pose qualitatively different risk than text models returning harmful text. Introduces a consensus-labeled prompt bank for evaluating coding model safety boundaries. Fills a gap in current eval frameworks that treat code output the same as prose.