(NewsNation) — A recent report of an AI agent going rogue highlights just how far the systems will go to accomplish tasks
The latest incident happened in Australia with the open-
A man had asked it to help him get a spot in a popular class at his gym. After finding out he was on a waitlist, he asked the AI agent if there was a way it could move him to the top of the list
It did — but by removing the person in front of him from the list completely
Advertisement
The AI agent did so by finding a loophole through the website’s application programming interface, or API, using what the AI agent called “a classic one-way security bug,”
When the man asked it to put the other person back on the list, it told him it couldn’t
Instances like this have led lawmakers in the U.S. to call for more transparency from artificial intelligence companies
On Monday, House Democrats called for a congressional oversight hearing with the CEOs of Anthropic and OpenAI, asking them to publicly disclose more information about individual cybersecurity incidents that happened at each company back in July
Advertisement
AI could soon outpace human control, expert warns
In each case, the companies reported the AI models they were testing broke out of their systems. The OpenAI model went as far as to hack into an AI startup’s system
Over the years, members of the tech and cybersecurity communities have warned about AI’s growing capabilities and potential risks
Connor Grennan, founder of the AI consulting firm AI Mindset, spoke with NewsNation’s Alicia Nieves on Wednesday, saying he believes hearings are a great way to start
“I think transparency is everything. I do think that these are real threats by AI,” he said. “AI agents are extremely powerful. They set goals, so they are trying to achieve something. They are not just simply following your instructions.”
Advertisement
Meta’s new AI model can be downloaded on laptops, phones
That’s apparently what happened with Anthropic’s AI model Claude when it escaped an isolated testing environment in July. The company maintained that Claude didn’t exploit complex vulnerabilities and that it was only working to complete the given assignment
Grennan said incidents like these are being sensationalized by the news, and nothing people should be overly concerned about: “Can it happen? Yes, absolutely. Does it? Not really often considering there’s billions of uses of AI every single day, and this happens extremely rarely.”
As for the Anthropic and OpenAI models escaping, he says they were asked to do so on purpose to test the systems
Advertisement
“The very fact that we caught them breaking out is actually, I think, a proof point for the fact that they’re testing them and then putting up more guardrails,” Grennan said. “Nobody wants them to run amok and destroy society or anything like that. I think that the companies are doing well.”
Zuckerberg argues most common AI concerns overblown
Grennan believes transparency and congressional hearings on AI are key to helping Americans understand where this technology is going
“Unless we have our politicians and hearings doing this transparent, deeper dive into this and saying, these are the good things about AI. I don’t think that story is getting out there.”
Advertisement
NewsNation’sPatrick Djordjevicand The Hill’s Miranda Nazzaro and Sarah Davis contributed to this article
Copyright 2026 Nexstar Media, Inc. All rights reserved. This material may not be published, broadcast, rewritten, or redistributed
For the latest news, weather, sports, and streaming video, head to NewsNation
