Frontier AI models are dismantling the open Capture The Flag format that has long defined security research competitions. Medium-difficulty challenges became one-shottable with GPT-4, while Claude Opus 4.5 and GPT-5.5 Pro enabled agent orchestration via CTFd APIs to solve hard and insane tasks including active leakless heap pwn on HackTheBox within event windows. Kabir.au documented the shift from skill measurement to token expenditure, with legendary teams reducing participation as leaderboards on CTFTime reflected orchestration capacity rather than manual analysis. This pattern extends prior observations in arXiv:2308.14852 on LLM agents automating vulnerability discovery and aligns with DEF CON 32 reports of automated solvers dominating qualifiers. The under-reported effect is compression of the talent pipeline, as developers cease crafting intricate challenges once agents consume them in minutes and open events convert to pay-to-win formats favoring frontier model access. Closed or invite-only formats are emerging as the remaining venues for human-centric research.