Aim Intelligence's Key AI Security Tech Paper Accepted for ACL 2025

AI security company AimIntelligence has solidified its international standing in AI safety research through innovations in LLM (Large Language Model) attack technology. Notably, the paper "One-Shot is Enough: Consolidating Multi-Turn Attacks into Efficient Single-Turn Prompts for LLMs," accepted to the main conference of ACL 2025, the most prestigious academic conference in natural language processing, garnered attention by proposing the 'M2S (Multi-turn-to-Single-turn)' framework, which compresses multi-turn attacks into single-turn prompts. The M2S framework utilized three unique strategies – hyphenation, numeration, and pythonization – demonstrating its ability to effectively attack LLMs without the need for traditional repetitive dialogue. The research results were astonishing. It achieved an attack success rate of up to 95.9% on the Mistral-7B model, and for GPT-4o, it even showed up to 17.5% higher effectiveness than existing multi-turn attacks. It is noteworthy that these results were achieved with 70-80% fewer tokens. This suggests that simple single-turn prompts can be as powerful as, or even more powerful than, multi-turn attacks, revealing serious vulnerabilities in current LLM defense systems. KAIST researcher Kim Hyun-jun led this achievement during his AimIntelligence internship, together with co-first author Ha Jun-woo. CEO Yoo Sang-yoon emphasized that this research "provides important insights for red teaming and safety mechanism design, and is a case where our AI safety research capabilities have been internationally recognized." AimIntelligence is also proving itself a leader in domestic and international AI safety research with a series of achievements, including the acceptance of its 'SUDO (Screen-based Universal Detox2Tox Offense)' framework research to the ACL 2025 Industry Track, and its VLM harm evaluation system 'ELITE' to ICML 2025.