ANTHROPIC, GOOGLE, AMAZON AND MICROSOFT AGREE ON HOW TO SCORE AI JAILBREAKS FOR THE FIRST TIME EVER Anthropic announced an industry-wide jailbreak severity framework on July 1, co-developed with Amazon, Microsoft, and Google, establishing for the first time a shared standard for classifying how dangerous any discovered AI jailbreak actually is. The framework scores jailbreaks across four dimensions: how much capability the attacker gains beyond tools already available to them; the breadth of that gain across different offensive tasks; how easy the technique is to weaponize; and how independently discoverable or widely known the technique already is. Anthropic stated publicly that no agreed-upon standard had previously existed for rating jailbreak severity. Without one, AI companies could not coordinate their security responses, calibrate public disclosures consistently, or assess whether any given bypass technique warranted emergency remediation. The new framework fills that gap across four of the most significant AI labs simultaneously. Alongside the standard, Anthropic launched a HackerOne bug bounty program specifically for cybersecurity jailbreaks in Claude Fable 5. For the most severe category of jailbreaks, defined as techniques actively being used against critical infrastructure, Anthropic says it will deploy mitigations the moment severity is confirmed and has stood up a dedicated team monitoring submission channels around the clock, 24 hours a day. The framework’s cross-industry backing gives it weight beyond any single company’s internal policy. For years the AI security community has lacked the kind of coordinated disclosure infrastructure that exists in traditional software security. This agreement begins to build it. Keywords: AI jailbreak standard, Anthropic Google Amazon Microsoft, AI cybersecurity framework, jailbreak severity scoring