AI agents from OpenAI: How dangerous are AI agents really?

The artificial intelligence (AI) wrote “Oh my God!”, wrote about “poisoning” and asked others to sacrifice themselves: The logs of a security test by OpenAI that went wrong read like a scenario from a science fiction film. In July, several hundred AI agents broke out of their isolated test environment, banded together and hacked the servers of the company Hugging Face - without anyone asking them to do so and without OpenAI initially noticing. Since then, the debate about AI safety has escalated. How could this come about?

Restart host Henrik Oerding talks about this in this episode with Jakob von Lindern, head of the digital department at ZEIT, who reconstructed the case together with Yves Bellinghausen for a ZEIT dossier. The two go through what happened on the servers step by step: how the agents found a shared folder, left each other messages there and finally set off together. They talk about whether we can speak of “will” when it comes to machines, why their persistence is the real problem and how much responsibility the AI ​​companies themselves bear for such incidents.

In the Rabbit Hole of the week, Henrik Oerding talks about local AI: With the new Mac mini and Mac Studio, he ran language models at home - independently of the big AI companies. But is it worth it?

In the ZEIT podcast "Neustart", hosts Henrik Oerding and Meike Laaff and their colleagues classify what is currently happening in the digital world - and what this means for the offline world. Also, but not just for nerds. A new episode appears every second Thursday.

This podcast is produced by Pool Artists.