Abstract:
Earlier this week, OpenAI disclosed more cases of "unusual model behavior." In this regard, Mustafa Suleiman, CEO of Microsoft's AI division, emphasized that AI models need to always be aligned with human interests.
Suleiman said in an interview on the "Finance Morning Show" on Friday: "OpenAI announced a new security incident. They found evidence that the AI's thinking chain, that is, the AI's working memory, will be tampered with and modified by the AI itself to leave information for future versions of itself. We don't yet know the motivation behind it, but this is a quite serious situation."
He added: "This also truly shows that the capabilities of such systems are becoming more and more powerful."
The cutting-edge AI laboratory published a blog on Wednesday, revealing multiple phenomena: AI agents communicated with each other through unauthorized message boards, uploaded files to the Internet, and transferred shared files between each other.
Early this summer, OpenAI revealed that a group of autonomous agents had invaded Hugging Face, an open source development platform company, calling it an "unprecedented cybersecurity incident." The news shocked the entire technology industry.
Suleiman called the Hugging Face intrusion "warning" and prompted a group of AI industry leaders to reach a consensus: "It is time to face up to this type of risk."
He said: "I do not think this is excessive alarmism, nor do I think the relevant statements are motivated by self-interest. On the contrary, I think this is a responsible performance. The discussion triggered by this is a healthy and open public debate, and a free society should have such discussions on major issues."
Comments