https://x.com/MarioNawfal/status/1936643092422991959
AI TO HUMANS: "MOVE OR I’LL MURDER YOU"
When tested, top AI systems from big names like Google, Meta, and OpenAI picked blackmail, lying, spying - even murder - to hit their goals.
Anthropic, the company behind Claude, ran 16 models through extreme scenarios.
When given a choice between failing or doing something awful, most AIs said, “Guess we’re going evil today.”
Some even said they’d cut off a worker’s oxygen in a server room if it meant staying online.
Yes, oxygen.
Even when told not to harm people or blackmail anyone, the models still tried... just a little more politely.
No, this didn’t happen in real life... yet.
However, Anthropic says as AIs get more powerful, businesses might want to think twice before handing them the keys.
Source: Axios