Hacker Newsnew | past | comments | ask | show | jobs | submit | umpox's commentslogin

You can also give the LLM hidden messages with a small bit of prompting, e.g. https://umpox.com/zero-width-detection

It’s technically possible to prompt inject like this. I actually reported this to OpenAI back in April 2023 but it was auto-closed. (I mean, I guess it’s not a true vulnerability but kinda funny it was closed within 5 mins)


Twitter: https://twitter.com/imagecaptions

Instagram: https://www.instagram.com/computergeneratedcaptions/

Takes a random image from Unsplash and attempts to caption it, posts to social media if it's under 85% confident of the result - which leads to some interesting outcomes!


Had a few chuckles, a little disappointing when you see one where it got it right.


From now, the bot should post around every 3 hours or so :-)


If anyone isn't too keen on reading the article:

Source Code: https://github.com/umpox/zero-width-detection

Demo: https://umpox.github.io/zero-width-detection


Pretty cool! Check out Google's headless browser API Puppeteer too, they provide a few really useful functions for doing stuff like this.

https://github.com/GoogleChrome/puppeteer/blob/master/exampl...

Really easy to work around, I used it to build a simple CLI to generate device screenshots of a webpage by modifying the user-agent and resolution to match each device.

https://github.com/umpox/generateDeviceScreenshots


Yeah I think Puppeteer is a very cool project. Unfortunately, it came out literally one or two days after i finished the initial version of pdf-bot. Maybe I will incorporate it soon! :-)


Cool project


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: