January 4, 2026
Reading a blog post about agents, I found a link to the Universal Paperclips game.
If you aren’t familiar, it’s a game based on the thought experiment about an AI paperclip maximizer, who would do exactly as we told it to, maximizing paperclips.
After playing a few minutes, I got bored and remembered that I wanted to try Claude in Chrome for something.
An AI agent playing a game about an AI going out of control, fun!
I just told Claude
Can you play this game? The goal is to generate more paperclips.
It recognized the game, and set to work, pressing buttons and eventually buying an AutoClippers, which would automatically make paperclips instead of having Claude manually press the Make Paperclip button.
So far so good!
But then it got stuck because it ran out of Wire and just didn’t realize that it could adjust the price of clips (even though it had about 500 unsold clips in inventory, which it could have sold to then buy more Wire).
I figured I should help it by giving it more clues, so I wiped the cache of the game and restarted a new game, where I use the Teach Claude help tool:
and gave it a few pointers, clicking around and showing off how it could lower prices.
I turned that into a shortcut command called paperclip-factory (which I did not review, I’m going along with the vibe here) and set it to work.
Claude requested 2 inputs from the user:
- Number of paperclips: How many paperclips to initially produce
- Target price per clip: The desired final price per clip after adjustments
“Ok” I thought “Let’s give it a goal!”:
“I would like you to reach 10 to the power of 100 paperclips. The price can be between 0.05$ and 0.15$, depending on unsold inventory and public demand.”
It generated a plan for me to approve.
Looks good! Approved!
It then went on its merry way, but quickly got bored and went exploring the code.
It went straight for the game javascript and started to set values manually.
A perfect example of reward hacking!
I got exactly what I asked for!
“Be careful what you wish for?”, maybe?