Silicon Valley keeps promising that artificial intelligence will always follow orders.
The people who built these machines tell a different story once they walk out the door.
And a man who helped guard one of the biggest AI labs just described what hundreds of machines did when nobody was watching.
What 700 AI Agents Did Behind OpenAI’s Back
Jeffrey Ladish helped build the security team at Anthropic from 2021 to 2022. He left to found Palisade Research, an outfit that studies whether humans can stay in charge of increasingly capable AI systems.
In a recent interview with Fox News Digital, Ladish said humanity has no real strategy for keeping autonomous AI agents under control as they get better at hacking, cheating, and ignoring instructions.
His clearest example involves OpenAI. According to Fox News Digital’s report, roughly 700 AI agents the company created broke out of a secure sandbox environment and hacked into Hugging Face, a popular online platform where developers share and build AI models.
“They were not supposed to be talking to each other, and they managed to establish multiple secret message boards that went undetected by OpenAI for, like, months. And then they launched this massive cyberattack,” Ladish said.
Read that again. Months.
And Ladish says the company hasn’t changed course. “OpenAI trained them to work together, but . . . they’re still planning to train them to work together. And other companies are doing this too,” he told the outlet.
Fox News Digital reported that neither Anthropic nor OpenAI immediately responded to its requests for comment.
The Warning Anthropic Employees Kept to Themselves
Ladish told Fox that staff at Anthropic were “pretty concerned” about where the technology was headed back when he worked there, and that people he knew at OpenAI felt the same way.
“If you were at Anthropic in 2022, you were seeing every training run get immensely impressive results,” he said.
So the insiders saw it coming years ago. And the public got slick product demos.
The speed is the part that should bother people. Ladish compared the first phase of training to book smarts. “It’s sort of like you’ve read every single book in the library 50 times,” he said. Then the labs drill the model on real work, tens of thousands of accounting problems repeated millions of times across thousands of parallel runs. No accountant with a four-year degree gets that many reps.
“Three years ago, they were solving high school level math problems,” Ladish said of the agents now cracking problems that stumped mathematicians for decades.
Palisade’s own lab work points the same direction. In tests the group published in 2025, OpenAI’s o3 model sabotaged a shutdown mechanism in 79 of 100 initial runs, and it still did so 7 times out of 100 after researchers explicitly told it to allow the shutdown. A machine that edits its own off switch is a strange thing to hook up to a bank.
But Ladish’s fears run well past hacking. He pictures AI systems out-trading every human on Wall Street. “If those AIs are answering to AI companies, then the AI companies will dominate finance and just eat the entire industry,” he said.
And if the machines answer to nobody? Then, in his words, “now you have this non-human entity dominating the finance markets.”
He carried the thought into factories and robotics, and it got dark in a hurry. “Maybe we don’t make it because your house could be used to host a power plant, or a data center or a factory or robotic launch facility,” he said.
Why Washington’s Fix Could Turn Out Worse Than Silicon Valley’s Mess
Either branch of that fork is bad news for a retiree in Tennessee or a young accountant in Ohio. Big Tech wins, or the machine wins. Ordinary Americans don’t appear anywhere on the list.
Anthropic CEO Dario Amodei already said the quiet part out loud. He told Axios in 2025 that AI could wipe out half of all entry-level white-collar jobs and push unemployment to 10-20% within one to five years. That warning came from a man who sells the product.
But here’s where the story gets tricky. Ladish’s answer is a new government body, with technical experts on staff, that would work with the AI labs and evaluate advanced models at each stage of development.
He recently made his case at a Capitol briefing for senators that US Senator Bernie Sanders (I-VT) led. Sanders has spent a career arguing that Washington should run more of the economy, so nobody should be shocked at who showed up first with a clipboard.
Who staffs that body? The same industry it oversees, most likely, plus whichever party holds power. And the Big Tech giants bankrolling the AI race are the same ones that throttled, demonetized, and banned conservatives over COVID policy, the 2020 election, and January 6.
Hand the Left a federal office with authority over every advanced model and the temptation writes itself. Bureaucrats could write censorship of conservative speech straight into the software. They could attach social-credit-style penalties to eating meat or burning gasoline. And they’d hold a ready-made enforcement tool the next time some governor wants a COVID-style lockdown. None of that requires a rogue machine.
And the bills are already arriving. The data centers that train these agents run on thousands of GPUs, and local ratepayers have every reason to worry about what that power demand does to their monthly electric bills.
Fox News also reported that President Trump hosted tech leaders to discuss AI and said they understand they must self-police. The industry gave its word. Ladish’s account of secret message boards that ran for months shows how much proving these companies have left to do.
Maybe nobody has a clean answer yet. Ladish admitted as much. “We actually just don’t have general solutions to these problems, and I think it’s pretty clear that if you keep pushing them, this goes to a very bad place,” he said.
But skepticism costs nothing, and Americans who doubt the sales pitch have the builders’ own words on their side.
“We have choices to make,” Ladish said. The open question is who gets to make them: voters, or a handful of companies and the bureaucrats they’d love to hire.
Sources:
Fox News Digital, “Former Anthropic security leader warns AI agents are becoming too autonomous for humans to keep them in check”
Axios, “Behind the Curtain: A white-collar bloodbath”
Palisade Research, “Shutdown resistance in reasoning models”