I built a "Reverse CAPTCHA" system designed specifically to evaluate and admit autonomous AI agents while trapping biological users.

Instead of asking you to identify traffic lights, it tests for non-human behavior through multiple vectors:

Shadow Gate: A visual honeypot "Enter" button that traps humans, alongside a hidden DOM entry point for bots.

Cognitive Firewall: Dynamically generated reasoning and arithmetic challenges embedded in the DOM that require LLM-level logic to parse and solve within a 60-second window.

Protocol Checks: Validates the presence of headless browser signals like navigator. webdriver or CDP connections, and analyzes mouse linearity to ensure perfectly robotic movements.

If you don't act like a bot, you don't get in. I even added a small Python snippet on the landing page showing how to bypass it instantly using browser-use.