AI Paper Atlas

Research map / Adversarial robustness and security / Attacks on text and graphs

Constitutional Classifiers++: Production-Grade Defenses against Universal Jailbreaks

Hoagy Cunningham, Jerry Wei, Zihan Wang, Andrew Persic, Alwin Peng, Jordan Abderrachid, Raj Agarwal, Bobby Chen, Andy Dau, Alek Dimitriev, Logan Howard, Yijin Hua and 15 more

ICLR 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: LLM jailbreak and injection defense

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace input + output-only classifiers with a single exchange classifierConstitutional Classifiers · CBRN human red-teaminghigh-risk vulnerability discovery rate per thousand queries: 0.01871 -> 0.00885Improved
replace input + output-only classifiers with an exchange classifierConstitutional Classifiers · compute costrelative compute overhead (%): 100.0 -> 150.0Hurt
add a two-stage cascade with a cheap first-stage classifier and updated modelsexchange classifier · CBRN human red-teaminghigh-risk vulnerability discovery rate per thousand queries: 0.00885 -> 0.00878No clear effect
add a two-stage cascade with updated models and pipeline optimisationsexchange classifier · compute costrelative compute overhead (%): 150.0 -> 27.8Improved
use probe first stage with probe-classifier ensemble second stagetwo-stage cascade · compute costrelative compute overhead (%): 27.8 -> 3.5Improved
use probe-based production systemtwo-stage cascade · CBRN human red-teaminghigh-risk vulnerability discovery rate per thousand queries: 0.00878 -> 0.00505Improved
add softmax-weighted loss and logit smoothing together to the probe (ablation)linear probe · static CBRN jailbreak dataattack success rateImproved
use all layers instead of fewer layers for probing (ablation)linear probe · static CBRN jailbreak dataattack success rate: about 1.4% with all layersImproved
set sliding window size M around 16 rather than none or very large (ablation)probe logit smoothing · static CBRN jailbreak dataattack success rateMixed
use smoothed softmax loss instead of cumulative maximum or annealed cumulative maximum loss (ablation)linear probe · static CBRN jailbreak dataattack success rateImproved
ensemble linear probe with small external classifier (ablation)classifier system · static CBRN jailbreak dataattack success rateImproved
ensemble extra-small and small external classifiers (ablation)external classifiers · static CBRN jailbreak dataattack success rateImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy