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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace input + output-only classifiers with a single exchange classifier | Constitutional Classifiers · CBRN human red-teaming | high-risk vulnerability discovery rate per thousand queries: 0.01871 -> 0.00885 | Improved |
| replace input + output-only classifiers with an exchange classifier | Constitutional Classifiers · compute cost | relative compute overhead (%): 100.0 -> 150.0 | Hurt |
| add a two-stage cascade with a cheap first-stage classifier and updated models | exchange classifier · CBRN human red-teaming | high-risk vulnerability discovery rate per thousand queries: 0.00885 -> 0.00878 | No clear effect |
| add a two-stage cascade with updated models and pipeline optimisations | exchange classifier · compute cost | relative compute overhead (%): 150.0 -> 27.8 | Improved |
| use probe first stage with probe-classifier ensemble second stage | two-stage cascade · compute cost | relative compute overhead (%): 27.8 -> 3.5 | Improved |
| use probe-based production system | two-stage cascade · CBRN human red-teaming | high-risk vulnerability discovery rate per thousand queries: 0.00878 -> 0.00505 | Improved |
| add softmax-weighted loss and logit smoothing together to the probe (ablation) | linear probe · static CBRN jailbreak data | attack success rate | Improved |
| use all layers instead of fewer layers for probing (ablation) | linear probe · static CBRN jailbreak data | attack success rate: about 1.4% with all layers | Improved |
| set sliding window size M around 16 rather than none or very large (ablation) | probe logit smoothing · static CBRN jailbreak data | attack success rate | Mixed |
| use smoothed softmax loss instead of cumulative maximum or annealed cumulative maximum loss (ablation) | linear probe · static CBRN jailbreak data | attack success rate | Improved |
| ensemble linear probe with small external classifier (ablation) | classifier system · static CBRN jailbreak data | attack success rate | Improved |
| ensemble extra-small and small external classifiers (ablation) | external classifiers · static CBRN jailbreak data | attack success rate | Improved |
Closest papers by meaning
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.88)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.88)
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.88)
- One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs (ICLR 2025 · similarity 0.87)
- Best-of-N Jailbreaking (NeurIPS 2025 · similarity 0.87)
- h4rm3l: A Language for Composable Jailbreak Attack Synthesis (ICLR 2025 · similarity 0.87)
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.
