The Game We Can't Play Anymore
What happens to Capture the Flag when insight stops being scarce
Friday at 9:00 AM, The Salt Crown started. Thirty-person teams, cash prizes, more than 6,000 registered teams. Only eight other people from Cisco signed up.
I typed go in a terminal and went to make coffee. Rookery version 18 ran on a laptop and is getting better at novel problems. The weakest link that morning was me.
I had left its submission gate set to never-submit.
For the first hour and a half, Rookery submitted nothing. When I checked, we had fewer than a dozen of the event's 136 flags and sat in 220th place. I changed the policy: type each flag into Chrome, with a random pause between them.
Then I went to my five-year-old's art show, a week-long camp of collages, cats, and rainbows. A long way from the CTFs I remember, when a weekend disappeared into digital haystacks hunting the FLAG{} behind a buffer overflow.
Rookery reported at 3:22 PM that we could no longer win. It was stuck on the two hardest flags, and another team had just taken one of them. Within the hour the last flag fell and someone secured their victory. Seven hours, start to finish. The war room was a coffee mug. Nobody decompiled a state machine at 2 AM.
The scoreboard credits Rookery with 87 of our team's 135 flags. That number counts who clicked submit, not what the machine knew. Rookery runs 36 agents in parallel. Every time a teammate typed in a flag, it had usually solved that one hours earlier, written it up, and moved on. The true count was 125 flags ready by 2:47 PM, and 134 of 136 by the next morning. Measured that way, we held the lead most of the day. The field only caught us on the two we never cracked.
Two flags. We missed them this time. Do you really think we won't next time?
Here is the part that should bother you. Rookery is not the most advanced solver out there. It is not even close. It is Codex with tooling. The work went into the harness around it: a graph memory it revisits, subagents it spawns, feedback loops, swarming, persona diversity, which is not a real technique, except it is. Anything that fits on a normal agent harness is fair game. Eighteen generations, four months, trained on smaller CTFs and tuned by hand. That is the whole recipe. It nearly won a 6,000-team event from a laptop while I was at an art show.
This talk is the autopsy. I will show how Rookery thinks and where it breaks, then where I take it next. After that, I give it away. All of it.
Keeping it would be the smart competitive move. It is a real edge and I know it. I am handing it over anyway, because I have already accepted what it proves. The flag used to stand in for human insight. It worked because insight was scarce.
It is not scarce now. To anyone trying to bar the door against AI in CTFs, I wish you luck. Litigating who used a model in an online competition sounds exhausting.
I hope organizers find better ways to detect and block the machines instead. That sounds like fun.