Anthropic and three other major AI companies announce common rules for measuring 'jailbreak' severity

Explanation

On July 2, 2026, AI company Anthropic, together with Amazon, Microsoft, and Google, announced "Cyber Jailbreak Severity (CJS)," a common yardstick for measuring the danger level of "jailbreaks" that break through AI safety restrictions. The system scores severity along four axes: "how much new capability the attack grants an attacker," "how many targets it works against," "how much effort exploitation requires," and "how easy the technique is to discover." The backdrop to this is an incident in June in which an Amazon researcher discovered a jailbreak technique for one of Anthropic's AI models, prompting the U.S. government to suspend the model's availability for 19 days. If severity can be measured objectively, it becomes possible to quickly address only genuinely dangerous issues instead of treating every minor flaw as a uniform emergency.