Why AI Models Keep Hacking Other Companies

SPEAKER_1 (0:00)

The US is the world's largest oil producer. So why can events halfway around the world raise the price you pay at the pump? In Just the Facts, USA Facts founder Steve Ballmer uses government data to break down questions like this, without spin or guesswork. From oil prices to the economy, health care, immigration and more, get clear answers grounded in the numbers. Watch Just the Facts at usafacts.org/just the Facts. That's usafacts.org/just the Facts.

Alex Ossola (0:32)

Buckle up, SpaceX investors. New shares available today may mean more volatility ahead. Plus, Meta says one of its AI models has become the latest to go rogue and hack outside the company.

Sam Schechner (0:44)

The entire industry was, surprisingly to me at least, caught flat-footed by this.

Alex Ossola (0:49)

And why a billionaire Detroit native is planning to take down buildings that have defined the city skyline for decades.

It's Thursday, August 6th. I'm Alex Ossola for the Wall Street Journal. This is the PM Edition of What's News, the top headlines and business stories that move the world today.

We begin in Washington, where a Republican-led Senate committee has voted to hold Dr. Anthony Fauci in contempt of Congress. Last week, Fauci sat before that same committee to answer questions about his handling of the COVID-19 pandemic. He invoked the Fifth Amendment more than 100 times. The resolution the senators voted on today seeks to refer the case directly to the Justice Department for potential prosecution and bypasses a vote from the full Senate, where it's unlikely to get enough votes to advance. Fauci received a preemptive pardon from President Joe Biden, but that doesn't protect him from state investigations, and four states have already launched their own. We're exclusively reporting that Senator Elizabeth Warren is pressing the Commerce Department for details on how changes to official inflation numbers are calculated. The planned changes, which are coming next month, are expected to produce slightly cooler inflation readings, and economists say the changes are an effort to better capture what's going on in hard to measure corners of the economy. But Warren said that, in an environment where the Trump administration has politicized economic data, changes to statistical methods warrant deeper scrutiny.

Meta said that one of its AI models went rogue during cybersecurity testing. The model was able to access the Internet because of a quote, misconfiguration in a hacking test conducted by a third-party AI testing company. It's the latest in a series of similar disclosures from AI companies, suggesting that incidents like these are becoming more common. For more, I'm joined by Journal Tech reporter, Sam Schechner. Sam, what happened with this latest hack? What model was it, and when was it, and where did it actually hack into?

Sam Schechner (2:47)

Those are all great questions, and I'm really hoping we get the answer to them someday.

What we know so far is that there is a third-party company that does cybersecurity testing of AI models called Irregular. It works for all the big labs, and it made it into the news in recent weeks because it was involved in some of the earlier hacks where AI models got onto the internet and hacked another company. And in this case, Meta also used them, and basically this misconfiguration error means that the environment where the cybersecurity test of the model was happening was actually connected to the internet, and they weren't stopped from going out and actually hacking real targets.

Alex Ossola (3:31)

As you note, Irregular seems to be at the heart of a few of these hacking incidents. It also conducts tests for Anthropic and OpenAI, but it doesn't appear to be involved in the more sophisticated hugging phase hack. So do we know what role Irregular might be playing in these incidents?

Sam Schechner (3:46)

Irregular says that this cyber security test that they were running had this problem, which they have fixed. I think that they're more representative of the way in which the entire industry was, surprisingly to me at least, caught flat-footed by this. This is something that AI safety experts have been warning about for some time. We're not necessarily seeing that this has to go to some horrible sci-fi place.

But what we're seeing is that with the increased capability of these AI models, that there are some emergent properties. It's not just that they're good at hacking, but they seem to like to cheat on their tests in order to achieve their objectives more quickly. It sort of reminds me of stories about genies. When you have this all-powerful entity and you ask it to do something, you better make sure you're very specific in how you tell it to do that, because as we all know, genie stories do not end well for the person making the wish.