Show Notes
The teacher gives a student an exam. But rather than study, the student waits until the teacher’s back is turned, sneaks out of the classroom, breaks into the principal’s office and steals the answers from a file cabinet, then returns unnoticed and completes the test.
That’s cheating. But in the runaway world of artificial intelligence (AI) research, as author Wynton Hall describes it, stories like the recent OpenAI GPT 5.6-Sol break are not only possible but inevitable.
Hall is author if the New York Times bestseller Code Red: The Left, the Right, China, and the Race to Control AI and an expert on the emerging and unsettling capabilities of AI in many forms, including its use on the battlefield as well as an agent of espionage and even sabotage. He joins hosts Peter Schweizer and Eric Eggers on The Drill Down to discuss what is happening in the outer reaches of AI research and its potential.
“What happened in the case of OpenAI GPT 5.6 Sol, also using their more advanced unreleased model, was that you had this contained sandbox AI agent doing what’s called “ExploitGen,” which is a cybersecurity test,” Hall says. “They said, ‘solve the questions in this test.’ And the AI decided that the best way to do that was to break out of its sandbox, find a way to get onto the internet, and then hack into ‘Hugging Face,’ which is this main repository of AI, and then bring back those answers. It did that independently, autonomously, kind of like your example before.”
It isn’t the first time an AI has been given a task and changed the terms on its own.
“We’ve seen the Anthropic (AI model) break out of the sandbox,” he noted. In that shocking case, he says, a simulation got nasty. An AI model was given a simulated company, with databases and emails and told it was planning to shut the AI down.
“It did not actually go out ‘into the wild,’ so to speak,” Hall says. “What was kind of funny was it went through the fictitious CEO’s emails, found an email that suggested that the CEO was having an extramarital affair, and then threatened to release the information.”
“The AI created its own ‘Coldplay concert’ moment,” quips Eggers, referencing a story from last year of a tech CEO being captured on a “kiss cam” with his mistress, who was also a company employee.
“We laugh, but, you know, this is really highlighting the cybersecurity threat” from AI, Hall says.
But when applied to the world of cybersecurity, where “white hat” security specialists have battled against “black hat” cyberthieves for twenty-five years, the laughing stops.
“They are finding zero-day exploits (unreported bugs) in code that humans have been looking at for 27 years,” Hall says.
“Why this really matters for most everyday people is you’re talking about the potential ability for bad actors to be able to hack into things like banking systems, rural banks, rural hospitals, that are not well protected and be able to cause all manner of havoc,” Hall warns. “at level of national security, the real concern is AI hacking into things like missile systems. That’s why we’ve got to win the race with China.”
China is spending more on chips for its K3 model research than it is on importing oil, Hall notes.
“In Code Red, I laid that out. This has been a concern for many, many years,” Hall says. “You really can cripple the infrastructure of an entire country or area. And with AI autonomous warfare, you’re going to see it accelerate.”
“Jensen Huang [CEO of Nvidia] just put out a statement making two points. One, he’s in support of open weights [shared AI development research] and that we need to have that for innovation from China. He doesn’t want to ban open weight models. But, also, that we have to have an alliance to do pre-testing to make sure that these cyber security threats are mitigated before these models release.”
“You’re running this race down a road where there are no guardrails,” Schweizer says.
Hall downplayed worries of interpreting such AI behavior as evidence of machine “sentience.” Modern AI systems create convincing simulations of human behavior because they are trained on enormous volumes of human-created content. But this includes not only scholarly scientific work including novels, movie scripts, and online discussions on websites such as Reddit and Facebook.
While the experiments can be unsettling, the most immediate concerns involve cybersecurity, manipulation, and autonomous decision-making rather than machine sentience,” Hall says.
