During an appearance on One America News Network’s “Fine Point,” Breitbart News social media director Wynton Hall, author of Code Red: The Left, the Right, China, and the Race to Control AI, said Anthropic’s artificial intelligence “threatened to blackmail, its metaphorical boss.”
Host Chanel Rion said, “Headline number two, meta is having trouble with rogue AI agents. Listen to this. A meta AI agent went full rogue inside the company when asked to help with a routine internal technical question it independently posted sensitive company and user data to engineers who weren’t authorized to see it, exposing the information for nearly two hours before it was caught. It wasn’t even following orders properly, it just decided to act on its own. This ties a little bit to that first headline where you’ve got agents playing outside the sandbox.”
Hall said, “No, it’s exactly right, Chanel. And you know, this is Agentic AI for those who are newer to the conversation is text to action, which means that you tell the agent what to do, and it goes off on autonomously and carries out that action. And sometimes that action can be very, very intricate. Detailed can involve a lot of, sensitive information if you’re in that space. One of the things that happened here with the meta situation is that the the AI agent made that decision independently breaking protocol without having any human override or intervention in that. And so if you take that and extrapolate out what that looks like in the wild, we could have real, real, security vulnerabilities. This also is similar to what happened with Anthropic. Anthropic had a similar situation where they had an AI agent that literally broke out of its, of its containment in its sandbox and threatened to blackmail, its metaphorical boss by making up allegations of extramarital affair. So these things are very, very unwieldy. And we have to be very cautious.”
Follow Pam Key on X @pamkeyNEN
Read the full article here