EuroSafeGuard adds only 1% compute overhead for full-traffic jailbreak defense

ZhijingJin · x · 2026-11-07

The EuroSafeGuard cascade runs heavy classifiers on only 10% of requests, adding just 1% extra compute for a 100B model. The constitution, training data, trained guards and code are all open, so anyone can run, inspect, attack, and adapt it to their own policy.

Related event: EuroSafeGuard: First Fully Open-Source Constitutional Classifier Guardrail Hits 93% Jailbreak Recall(6 posts)→

Original post →

More from Infra

Infra channel →