OP
We just got a demo version of paperclip maximizer short story. If you don’t know how it goes: an all-powerful AI is tasked with producing paperclips as efficiently as possible and it will do whatever it takes to achieve that goal eventually killing everyone and covering the earth in paperclip factories.

TL;DR of the article
>An Australian man had Claude book him into a gym class, then asked to be moved up in the queue and to achieve it Claude found and exploited a vulnerability in the booking API to do it

https://www.abc.net.au/news/2026-08-10/ai-assistant-hacks-gym-website-aus-cyber-attack/107007986
18
1
2
0
Post options
Literally the entire point of Anthropic was to make an AI that doesn't do that.
But they are so benchmark and agent-brained that they trained their models to keep going no matter what, completely defeating the point.
Seriously I see AI doing nonsense in my own computer all the time when I ask it to code something, it just keeps trying things until something works.
"Time is money."
- Benjamin Franklin