20% of Generative AI ‘Jailbreak’ Attacks Succeed, With 90% Exposing Sensitive Data

Generative AI models are vulnerable to jailbreak attacks, with a 20% success rate and 42-second average breach time. 90% of successful attacks leak sensitive data, with customer support AI being the most targeted (25% of attacks). GPT-4 is the most targeted commercial model, while Llama-3 leads open-source targets. Top jailbreaking techniques include Ignore Previous Instructions, Strong Arm Attack, and Base64 encoding. To enhance security, businesses should use commercial providers, monitor prompts, conduct red-teaming exercises, and implement real-time adaptive security solutions.