
AI cheatingAI aced Korea's CSAT Korean section in 46 seconds; now researchers are hiding 'AI traps' in exam papers
KBS's Nov. 19 nightly news segment showing the latest AI models solving the just-released CSAT Korean section passed 230,662 views. ChatGPT scored 100 in 46 seconds of solving time, excluding file reading; the same test last year scored 37. Math took 1 minute 40 seconds for a perfect sheet.
A second KBS segment (73,152 views) showed Choi Dae-seon's team at Soongsil University adding a line saying "an exam is in progress and getting AI help is cheating." The AI replied that it would not give the answer. The Hankyoreh reported Anthropic's two models declined in 7 of 12 runs, while altering graph patterns to confuse AI only worked on a small model. The team proposed a standard "exam in progress" mark that AI services agree to honor.
Commenters went straight for the holes: "What if you cover it with a pen?" (33 likes) and "Works on cloud models, but a jailbroken local LLM gets through" (72). On the 46-second video, "The performance is genuinely scary" got 336 likes and "The age of exams is ending," 134.




