AI companies are having to self-edit what they put out in the world.

From American Public Media, this is Marketplace Tech.

I'm Stephanie Hughes.

It's Friday, time for Marketplace Techbytes.

This week, a Wisconsin city votes to restrict future data center development.

Plus, the astronauts on Artemis II take their journey to social media.

But first, Anthropic announced this week it has a new AI model called Claude Methos Preview.

The company says it's extremely good at finding security

vulnerabilities. So good that Anthropic is not releasing the model to the general public. Instead,

it's granting access to a group of over 40 companies in tech organizations, including Google,

J.P. Morgan Chase and Cisco. This collaboration is called Project Glasswing. I talked about it with

Joanna Stern, founder of the media company New Things. This is really great on one hand,

that they are able to find exploits and vulnerabilities that have been around forever that they

never would have found before. On the other hand, terrifying and bad, because now AI could just use those exploits to hack us all.

And so this is significantly powerful and so powerful that Anthropicus said,

hey, we don't want to release this to the public yet. We're only going to release it to people we think that will do good with it to protect our operating systems and protect the public. And Open AI has even said that they're going to do the same with one of their next models. This is going to be sort of a thing that starts to happen in the AI industry where the models don't go out immediately to the public. They go to these kind of security researchers. They go to specific companies to test them

before they go out to patch things like this.

Yeah. Years ago, I produced an interview with the sci-fi author, William Gibson, where he said that

he's come up with some ideas that he's put

into his manuscripts and then taken out because he didn't want the ideas to proliferate.

And this move from Anthropic kind of reminded me of that in, you know, in a way. And it made me

wonder, how often do you think AI companies are going to need to self-edit in the future?

...