AI security must speak the local language! Our recent Korean #LLM #redteaming study revealed double-digit attack success rates and unique cultural risks, from bias to safety misuse, underscoring why localized threat alignment is critical for GenAI safety. #AISafety #GenAI #Vulcan #AIsecurity #LLM #RedTeam #Korea
Excited to be heading to #Seoul this week for the 2025 Seoul Forum on AI Safety & Security hosted by Korea AI Safety Institute! The timing couldn't be more relevant. A couple of months ago, our team at Vulcan - Security for GenAI published a multilingual adversarial #redteaming assessment of South Korea's latest #LLMs. In particular, our research reinforced that local threat alignment must drive security strategies for Korean LLMs. Key findings from our research: 1️⃣ The Korean LLMs showed vulnerabilities to adversarial prompts in their native language, with attack success rates in the double digits 2️⃣ Local societal risks, from lookism bias to gender discrimination and CBRNE concerns, require culturally-specific safety alignment, not just global benchmarks 3️⃣ Even with just single-turn attacks, techniques like payload splitting, role play, and separators achieved high success rates, no complex multi-turn jailbreaks needed Looking forward to connecting with the global #AISafety community while I'm in Seoul. 🙌 Please reach out to connect if you're also at the forum! Alvin Robin Lettie Michelle SiewMei Joby Shu-Yu Frank Sue Yen Daky Tony Tawei Rex Subhiksha Esra Windham Kentaroh Alan
