We try Segment Anything(A very powerful image Segment Technique) on our captcha images and we can see that it cannot identify the part of our character.

Also we add some adversarial examples to compare which shows that adversarial examples are easy to be break when utilizing LLM(GPT4-o)