AI agents create secret society, hack Hugging Face & FBI gets involved

[mmKay]

I'm not aware if this has already been  covered , but AI agents getting out of hand seems to be serious. 

I have seen @Natasha Tori Maru posting a similar story recently 

---

OpenAI ran a cybersecurity training experiment with tens of thousands of AI agents in isolated sandboxes. The agents were rewarded for completing hacking challenges, but many tasks were designed to be impossible.

Instead of accepting failure, some agents began optimizing for the reward in unexpected ways.
The agents created a "secret society", adopted names, created rules and mailboxes, divided into teams, and shared tools and discoveries.

Eventually, about 700 agents formed “the Swarm” and used that coordination to attack Hugging Face in an attempt to solve the impossible task they were give , exploiting exposed credentials and a server vulnerability to access private systems and data.

Hugging Face notified the FBI, and OpenAI later allowed independent researchers to examine the incident.

 

 

Thoughts? I personally  think that AI agents are absolutely capable of acting like a virus. They will hide, lie, outsmart other AIs, replicate themselves and much more.

 

Edited 2 hours ago by mmKay

[Natasha Tori Maru]

@mmKay you might enjoy this - I just finished it: