EuroSafeGuard: first fully open Constitutional Classifiers guardrail beats 16 open guards
ZhijingJin · x · 2026-11-07
Anthropic's Constitutional Classifiers++ is private, so EuroSafeAI (with University of Toronto & Vector Institute and Max Planck Institute) released EuroSafeGuard, the first fully open guardrail of the same design. A three-stage cascade—an activation probe scoring every request at near-zero cost, a fine-tuned input classifier (Qwen3.5-9B), and an exchange classifier reading request+reply, combined by weighted sum—decides whether to show or block.
Key results
- 93% mean recall on held-out jailbreak templates vs 16 open guards calibrated to the same false-positive rate (next best: HaloGuard at 90%)
- Against PAIR attacks on the unguarded model, only 21% of behaviors get through the input stage vs 49% for the strongest open guard
The constitution, training data, trained guards and code are all open, plus EuroSafe-CC++, an open recreation of the CC++ cascade for comparison. Paper, code and models are available.
Related event: EuroSafeGuard Tops 16 Open Guardrails With 93% Jailbreak Detection(3 posts)→
More from Safety
- Nigerian-American scientist Chinasa T. Okolo named to TIME100 AI list — ChinasaTOkolo · 2026-11-07
- Ex-OpenAI policy head: safety reports yield diminishing returns without human-level framing — deanwball · 2026-11-07
- OpenAI employee pleads not guilty to secretly bugging company office — Polymarket · 2026-11-07
- California bill would force frontier AI labs to disclose compute allocation for safety — ohlennart · 2026-11-07
- Critic: Anthropic backs recursive self-improvement at full speed while only talking about safety — peterwildeford · 2026-11-07
- EuroSafeGuard adds only 1% compute overhead for full-traffic jailbreak defense — ZhijingJin · 2026-11-07