Research reveals that harmful content can be elicited from LLMs through ordinary conversational interactions, without requiring technical expertise.

• Simple multi-turn prompting strategies can bypass safety measures
• The "Speak Easy" framework successfully jailbreaks various commercial LLMs
• Proposed "HarmScore" metric evaluates real-world actionability of harmful outputs
• Multilingual testing exposes vulnerabilities across language barriers

This research highlights critical security implications for AI deployment in consumer-facing applications, emphasizing the need for more robust safety mechanisms against conversational manipulation techniques.

Speak Easy: Eliciting Harmful Jailbreaks from LLMs with Simple Interactions