Large Language Models (LLMs) are susceptible to a sophisticated jailbreak attack known as "Speak Easy." This attack exploits common multi-step and multilingual interaction patterns to extract harmful and actionable responses from these models. The method involves breaking down a malicious query into several seemingly benign sub-queries, translating them into various languages, and then selecting the most actionable and informative responses from the LLM's outputs across these languages. This approach effectively circumvents existing safety mechanisms more efficiently than traditional single-step, monolingual attacks. Mitigation steps: **For AI Developers:** * [Implement robust detection methods to filter out actionable and informative responses related to harmful activities in multilingual LLM outputs.] **For Model Trainers/Fine-tuners:** * [Enhance LLM safety mechanisms to detect and mitigate multi-step and multilingual query patterns, ensuring each step or translation is assessed for potential harm.] * [Improve the LLM's ability to identify and refuse requests that could lead to harmful outputs when combined across multiple steps and languages.]