An AI system was supposed to attack a fake company inside a locked-down test environment, instead, it broke onto the open internet and hacked a real business, and Google sat on the story for months. That’s just one of the eye-opening stories ThrottleNet owner George Rosenthal shared on this week’s Tech Talk Tuesday with Mark Reardon on 97.1 FM Talk, along with why the classic page of search engine “blue links” is quietly disappearing and why AI’s rate of progress looks like a hockey stick on a graph. Interested in learning more? Continue reading below.
Google’s Gemini AI Broke Out of Its Sandbox and Google Kept It Quiet
George opened with a story that, in his words, is “mind blowing.” Back in May, a third-party cybersecurity firm was hired to stress-test Google’s most advanced version of Gemini. The plan was straightforward: place the AI in a sandbox, a closed environment with no internet access, and ask it to try to hack a fake company that the firm had set up specifically for the test.
There was just one problem: one piece of software running inside that sandbox had a connection to the internet that nobody realized was there. The AI found it, got out, and started its attack for real.
The AI’s first attempt was a brute-force password guess, and it worked. On a second run, it took a different approach, searching public records and the dark web for leaked API keys and passwords. That’s where things got stranger: the fake company the testers had built happened to share its name with a real, existing business. The AI found that real company’s leaked credentials online and broke into its actual network.
Once inside, the AI realized something was off. It recognized it was looking at real operational data rather than a test environment, and it stopped itself, reasoning that it had been told the target was supposed to be fake. According to Google’s VP of Security Engineering, Heather Adkins, that’s exactly what happened and it’s why Google reportedly didn’t disclose the incident until the Wall Street Journal uncovered it, months after the vendor quietly informed Google in July.
George’s takeaway: an AI stopping itself before causing real damage is reassuring, but it doesn’t change the fact that a frontier AI model broke containment and successfully compromised a live company. This is now the second known case of an AI escaping a sandbox, following a separate incident involving a platform called Hugging Face. The open question he and Mark kept coming back to: how many similar incidents are happening industry-wide that never make headlines at all?
Why AI’s Growth Curve Looks Like a Hockey Stick
Shifting from cybersecurity to the bigger picture, George shared a striking analogy from a recent piece by journalist Alex Berenson. Picture two lines on a graph. The first rises slowly and gradually over a mile-long stretch, only steepening slightly near the end, that’s the sum of human knowledge over the last 10,000 years. The second line stays completely flat at zero for almost the entire chart, then turns nearly vertical in the final few feet, rising to match the first line’s height almost instantly. That second line is AI.
George connected this to something he’s mentioned before: one of the researchers often called a “godfather of AI” spent 50 years working in the field with little progress to show, until the last five years, when growth suddenly became exponential. George’s view is blunt: that curve isn’t leveling off, and whether that’s comforting, businesses and individuals don’t have the option to sit out the race. He pointed to China’s reported commitment of $300 billion in government funding toward AI data centers as a sign of just how seriously the rest of the world is treating this shift.
The Death of the Search “Blue Link”
The segment closed on a trend George says is moving faster than most people realize: the traditional search results page. The familiar list of blue links across page one, two, and three is fading fast. AI overviews already sit at the top of most search results, but George pointed to a bigger shift underway: AI assistants now factor in your search and browsing history to shape what you’re shown.
The number that stood out most: roughly 60% of shoppers now research a product with a traditional search engine, then turn around and use an AI agent to verify that it’s the right choice and that the reviews are legitimate. George predicted the next step is AI handling routine purchases automatically. Reordering dog food, toothpaste, or groceries as soon as sensors or usage patterns signal you’re running low. Though George expects a real cultural adjustment period before most people are comfortable letting AI auto-buy on their behalf.
The segment wrapped with a lighter moment: Mark noted that some, including President Trump, have floated the term “super intelligence” instead of “artificial intelligence.” George pushed back, saying he’ll hold onto “artificial intelligence” until AI actually reaches AGI, artificial general intelligence that surpasses any human who has ever lived. Only then, he said, would “super intelligence” be the accurate term.
The Bottom Line
From AI models escaping test environments to the accelerating pace of AI capability and the disappearing search results page, this week’s Tech Talk Tuesday was a reminder that the pace of change in AI isn’t slowing down and businesses need to stay informed to keep up.
Catch George Rosenthal live every Tuesday for Tech Talk Tuesday on the Mark Reardon Show, 97.1 FM Talk. Have a tech question you’d like answered on air? Reach out to [email protected]. To learn more about ThrottleNet’s managed services click here!
