Reproduce “Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hoagy Cunningham, Jerry Wei, Zihan Wang, Andrew Persic, Alwin Peng, Jordan Abderrachid, Raj Agarwal, Bobby Chen and 21 more
Anthropic researchers enhanced Constitutional Classifiers to provide production-grade defenses against universal jailbreaks, achieving superior robustness, a 40x reduction in computational cost, and a 0.05% refusal rate on production traffic. The system integrates context-aware exchange classifiers, a two-stage cascade for efficiency, and model-internal linear activation probes to effectively counter evolving adversarial attacks.
Key insights
- Context-aware evaluation via exchange classifiers is critical for defending against advanced adversarial techniques like reconstruction and output obfuscation.
- A two-stage classifier cascade significantly improves computational efficiency while maintaining high robustness against jailbreaks.
- Linear activation probes, by utilizing LLM internal activations, provide a highly efficient and complementary safety signal, enhancing ensemble performance and reducing costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.