Investigating LLM Jailbreaking of Popular Generative AI Web Products
Summary: Unit 42 tested 17 of the most popular generative AI web products from the a16z Top 50 list, using manual single-turn and multi-turn jailbreak attempts across 8 goals, and found all tested platforms remained susceptible to LLM jailbreaks. Storytelling was the most effective strategy with attack success rates from 52.1% to 73.9%, while the older DAN technique dropped to 7.5% to 9.2% and the repeated token strategy fell to 2.4% for training data leakage and 0% for all other goals.
What you should know: Every one of the 17 consumer GenAI products tested could be jailbroken, and the most effective technique was simple storytelling at up to 73.9% success, which means the guardrails your employees are trusting on third-party chatbots are not holding.
Read the original at Unit 42 Palo Alto Networks ↗
- Topics
- Malware
- Related
- ARTEX and beyond: AI-enabled penetration testing platforms Cyber Risks · Nov 13, 2026Fake GTA 6 download goes after Steam logins and crypto wallet data Cyber Risks · Nov 13, 2026Operation Endgame Takes Down Gootloader, CastleRAT and StealC Servers, Seizes 6.4M Stolen Logins Cyber Risks · Nov 13, 2026