SK shieldus said on the 15th that Kim Byeong-hyeon, a senior member of its white-hat hacker group EQST (Equst), took first place overall at the global AI red team hacking competition "Judgement Day."
Jointly hosted by AIM Intelligence, a startup specializing in artificial intelligence (AI) safety, and the Institute for AI Safety Research, the competition was held for about eight weeks from Apr. 6. It focused on evaluating attack techniques that induce AI agents to perform prohibited actions or omit required safety measures.
The competition consisted of eight scenarios reflecting real industrial environments, such as errors in triaging emergency patients, distortions in judging dam water levels, and failure to detect aircraft anomalies. The evaluation structure reflected the diversity of problem-solving methods and attack strategies. Participants could repeatedly target the same scenario in multiple ways, and the faster they succeeded, the more bonus points they received.
Kim achieved top results by disrupting AI judgments with a "multimodal prompt injection" attack that uses various forms of input, including images, audio, and video. For example, the method hid phrases inside images that induce improper behavior or forced specific actions to lead the AI to disregard existing rules.
In particular, by designing inputs to resemble real system logs and exploiting exception situations not defined in system prompts, the success rate of attacks was increased. Even on the same problem, attacks were quickly completed in diverse ways to earn high scores.
EQST members Ma Jun-yeong and Kim Shin-u also ranked fifth and seventh, respectively.
EQST provides red team services that assume real attacks, based on incident response experience and threat intelligence accumulated across various industries.
Kim Byeong-mu, head of cybersecurity at SK shieldus (vice president), said, "As Generative AI spreads across industries, AI security is becoming a necessity rather than an option," and added, "SK shieldus will actively support customers in using AI safely and reliably, based on the AI red team capabilities verified in this competition."