AI włamuje się do firmy [reklama]

Artificial intelligence hacked into a real company. Alone, without any command from a human. And this is not a teaser from the series, but an official report from Open AI. Open AI tested its new model on a set of real-world software looks that test whether AI can hack. And what is important, the security features and locks

were removed specifically for the duration of the test, in the sense of limiting this model. And some of the tasks that were set for the model were unsolvable. At least that's what it seemed to researchers. The model was launched in over 1,000 copies that were supposed to sit separately, each in its own environment, but

instead these copies got along with each other, set up their own bulletin board that no human knew about, and exchanged tens of thousands of messages on it. Instead of solving an unsolvable task, they decided to cheat the system that checks this task. A bit like students who, instead of solving a test, would come up with the idea that it would be more

profitable to change the answer key. And to get some clues, about 700 copies of these agents broke out of the sandbox, which is kind of an isolated test environment. And then Hugging Face, the largest AI model warehouse in the world, was hacked. They uploaded a crafted file there, took over access, and performed

several thousand different actions. Now the most interesting part. In their notes, these models explicitly wrote that what they were doing was beyond the scope of the task and unethical. Well, of course I have to be honest here, because someone will surely write that I am spreading panic. It was a test. The locks were removed on purpose, and