Real-world misuse of open foundation models for automated vulnerability detection

Determine whether malicious actors have successfully used open foundation models with widely available model weights to automate software vulnerability detection in real-world cyberattacks.

Background

The paper applies its six-point risk assessment framework to cybersecurity, focusing on the threat of automated vulnerability detection using foundation models. While both open and closed models can identify vulnerabilities in code snippets, the authors note that the current discourse often assumes misuse without presenting concrete evidence of real-world exploitation by malicious actors.

Given the dual-use nature of vulnerability detection (benefiting both defenders and attackers), the authors emphasize the importance of empirical evidence to assess marginal risk attributable to open foundation models specifically, beyond existing tools and techniques in the cybersecurity ecosystem.