OpenAI and FMF launch framework to grade cyber jailbreak risks

Jailbreaking AI models has turned into an internet sport, but security teams face a practical headache: not all bypasses are created equal, so which holes should they patch first?

To tackle that question, OpenAI has teamed up with the Frontier Model Forum (FMF) and its member companies to release a new evaluation framework. The goal is to pinpoint exactly when slipping past an AI system's guardrails creates serious cyber harm, helping teams prioritize their defenses.

Grading real-world damage

Instead of just treating every jailbreak as a binary failure, the new approach looks at the actual blast radius.

Sasha Baker, OpenAI's head of national security policy, noted that this framework offers a shared methodology to grade and compare cyber jailbreak severity. Crucially, it evaluates both the specific capabilities unlocked by a bypass and how readily an attacker can actually use them in practice. With a shared standard, the AI and security communities can build a clearer body of evidence to reinforce their guardrails.

Why it matters

We think this is a smart pivot away from the endless game of whack-a-mole that model safety has become.

Tricking a chatbot into generating buggy exploit scripts is common, but measuring whether that bypass actually lowers the barrier to a serious attack is what matters. Ranking jailbreaks by actual threat severity gives developers and enterprise defenders a practical roadmap for allocating resources.

The real test now is adoption. A framework only works if member firms and external researchers genuinely share threat data rather than keeping severe exploits under wraps.