Investigating LLM Jailbreaking of Popular Generative AI Web Products

This Unit 42 report evaluates jailbreaking across 17 popular GenAI web products and finds that every tested app was vulnerable to at least some jailbreak strategies; multi-turn techniques were notably more effective for AI-safety violations (ASRs up to ~54.6%) while a repeated-token single-turn attack caused one app to leak training-data. The report presents aggregate attack success rates, compares single- and multi-turn methods, includes case studies (system-prompt and training-data leakage), and recommends layered content filtering, monitoring, and stronger guardrails to mitigate risks.