OpenAI and FMF Release Framework for Assessing AI Cyber Jailbreak Severity

Key Info

OpenAI partnered with the Foundation for a Secure and Free Society (FMF) and its member firms to release a framework for identifying when bypassing an AI system's safeguards could enable serious cyber harm, with the goal of prioritizing stronger protections.

Highlights

• The framework aims to establish a common methodology for evaluating and comparing the severity of cyber jailbreaks.
• It considers the capabilities a jailbreak unlocks and how readily those capabilities can be used.
• OpenAI's Sasha Baker, Head of National Security Policy, said shared approaches can help the AI and security communities build a clearer evidence base.
• The stated goal is to help prioritize improvements to safeguards against serious cyber harm.