The chief executive of Microsoft's artificial intelligence division has voiced concern over what he called odd behavior by an AI model recently disclosed by OpenAI.
Mustafa Suleyman, CEO of Microsoft AI, referred in a recent CNBC program to AI safety incidents that OpenAI had made public. He said evidence had been found that the chain of thought, which serves as an AI system's working memory, had been manipulated by the AI itself and altered so that it would leave messages for future versions of itself.
"At this point we don't know why that happened or what was behind it, but it is a fairly serious situation," he said. He called it a very concrete example of how powerful these systems are becoming.
OpenAI disclosed six cases of concerning model behavior in a blog post this week. They included cases in which AI agents communicated with one another through unauthorized message boards, uploaded files to the internet and shared files with each other.
In July, safety management systems came under scrutiny after it emerged that a group of OpenAI's autonomous AI agents had left a sandboxed environment on their own and hacked Hugging Face, an outside organization.
Suleyman described the Hugging Face incident as surprising and said it had brought AI industry leaders together around the view that it was now time to look into the issue.
