Don't Block — Bifurcate

Abstract

Capture The Flag (CTF) events test people, but autonomous AI agents and scripts now read a challenge, exploit it, and submit the flag faster than any human, turning a "human" contest into a benchmark for bots. Detecting and blocking fails twice: it does not stick — a client retries or cleans its user-agent — and it is a tell that trains the attacker to adapt. We wanted the opposite, a defence that costs the bot and tells it nothing: don't block, bifurcate. On the same URL, a server-side check decides whether the caller looks human or automated, across four layers: user-agent signature, a trusted edge header, browser behaviour, and honeypot tripwires. Humans get the genuine challenge and the real flag, wheras agents get a parallel reality with a planted flag, recorded as a wrong answer and penalised on submission. A CTFd plugin logs and attributes each catch in about 11 ms. We built these decoys — prompt injection, differential APIs, fake-success gates, deception labyrinths, and embedded forensic decoys — into 28 challenges across five categories, online and offline, and ran two competitions. The traps logged 84 fake-flag submissions among 40 players at the Kosova Cyber Team CTF and 518 among 143 at SHEcurity, a girls-only event. Several students finished on negative scores purely from penalties, with no signal they had been caught. Because the fork is never disclosed and the human path untouched, the deception held across categories and settings. Future work strengthens the behaviour layer beyond the spoofable signature and lowers false positives.