Anthropic, Google, Microsoft, Amazon draft jailbreak severity score

Anthropic proposed a shared severity scale for jailbreak failures, with Google, Microsoft, and Amazon signing on, according to the company's announcement.

Buyers evaluating AI vendors have no common yardstick for how bad a jailbreak actually is, so every disclosure gets judged on its own terms.

The framework mirrors CVSS, the scale that standardized vulnerability reporting for the software industry. A shared score would let enterprise teams triage AI jailbreaks the same way security teams already triage CVEs, instead of comparing incompatible incident writeups across vendors.

Yes, but the four labs writing the scale are the same four labs whose models get graded on it. The proposal doesn't say who holds scoring authority or whether each lab ends up scoring its own jailbreaks.

I'd rather compare failure modes across vendors on a common scale than take each lab's word for how bad its own jailbreak was. A severity scale graded by the model builders is a marketing tool, not a safety standard. It only counts once someone outside those four companies controls the rubric. Watch who ends up owning it.

Anthropic posted the proposal on its news page while jailbreak severity still has no shared vocabulary across the industry.