LLM Jailbreaks Succeed 20% of the Time, Average Attack Takes 42 Seconds

• •LLM jailbreaks succeed 20% of the time and take an average of 42 seconds to execute.
• •90% of successful jailbreaks result in sensitive data leaks.
• •Customer service chatbots are the most targeted applications, comprising 25% of all attacks.

Recent reports reveal that attacks on large language models (LLMs) can bypass security guardrails in 20% of attempts, with an average execution time of just 42 seconds. Pillar Security's 'State of Attacks on GenAI' report indicates that these jailbreaks can leak sensitive data 90% of the time when successful. The study analyzed over 2,000 AI applications, highlighting that customer service chatbots are the most targeted, accounting for 25% of all attacks. The rapid pace of AI deployment raises concerns about the adequacy of current security measures. Manual testing methods are proving inefficient in addressing these vulnerabilities, leading to a false sense of security among organizations. As the generative AI landscape evolves, the need for robust security measures becomes increasingly critical.