How Anthropic’s Jailbreak Challenge Put AI Safety Defenses to the Test

**Executive Summary:** Anthropic partnered with HackerOne for a week-long AI red teaming challenge against a demo of Claude 3.5 Sonnet to evaluate Constitutional Classifiers for CBRN-related safety bypasses; 339 participants produced over 300,000 chat interactions, several teams earned bounties for successful jailbreaks, and the findings highlighted attack techniques and areas for improving model defenses.