← Back to context

Comment by cadamsdotcom

3 days ago

Bulletproof solution: captcha where you drag a cartoon wire to one of several holes, captioned “for access, hack this phone system”

No agent will touch it!

“As a large language model, I don’t hack things”

Captcha: "Draw a human hand with the correct number of fingers"

AI agent: *intense sweating*

I actually have had some success with AI "red-teaming" against my systems to identify possible exploits.

What seems to be a better CAPTCHA, at least against non-Musk LLMs is to ask them to use profanities; they'll generally refuse even when you really insist.