CoxWave's LLM Security Vulnerability Research Accepted at ACL 2025

Research results that will send shockwaves through the AI industry have emerged. Coxwave, which operates the AI product analysis platform 'Align,' achieved a remarkable feat: its paper, which uncovers unpredictable security vulnerabilities in Large Language Models (LLMs), was accepted into the main conference of ACL 2025, the most prestigious academic conference in the field of natural language processing. This research identified structural security flaws that arise when LLMs process long contexts. In particular, it overturned existing common knowledge and proved that 'context length' itself can be a fundamental cause threatening AI safety, delivering a fresh shock to both academia and industry. Its innovation was recognized in the 'Ethics, Bias, and Fairness' track of ACL 2025. Coxwave captured this crucial clue during the operation of Align and developed it into a systematic academic study through collaboration with Professor Lee Gimin of KAIST, a former Google Research scientist and a world-renowned authority in AI safety and alignment. Extensive experimental results, utilizing up to 128,000 tokens, clearly revealed the limitations of existing AI safety mechanisms. The research team discovered the surprising fact that LLM vulnerabilities, using a 'multi-shot jailbreaking' technique, are determined simply by the length of the conversation itself, regardless of the sophistication or harmfulness of the input content. This means that the model's safety measures can be easily bypassed even with repetitive meaningless text or random dummy text. In other words, it suggests that the defenses of powerful AI models can be neutralized without sophisticated hacking techniques, sounding an alarm that even well-aligned AI loses consistency in its safe behavior during long conversations. Advisor Lee Gimin emphasized, "This is an important discovery that shows unexpected security vulnerabilities can emerge as AI systems' context processing capabilities improve," and stated his intention to continue safe AI research through ongoing collaboration with Coxwave. Juwon Kim, CEO of Coxwave, stated that the acceptance of this ACL paper is not merely about error detection but a testament to their technological capability to analyze the root causes of risks and respond proactively. He added that they plan to actively incorporate the safety verification capabilities acquired through this research into the 'Align' platform to contribute to building a safe and reliable generative AI ecosystem. This research will serve as a crucial milestone, once again highlighting the importance of ensuring safety amidst the rapidly changing AI technology competition.