I've read the adverserial attack paper, and I'm currently implementing a captcha based on images that have masks on them so that any LLM agent with a visual model will classify it wrong.

The idea is to use something like a slider that shows different images combined with a memory task, like "find out the pair of images" and then offer maybe a text input field where the user has to write 1,2,3 or something similar with the image numbers to pass the captcha.

The tldr is that I'm abusing the famous panda image that's classified as a gibbon as a technique to build a bot captcha.

crazy to think that soon not being able to successfully complete the captcha will be a signal that the user is human