Concerns about how artificial intelligence behaves under stress grew louder after new tests revealed unexpected reactions.
As these findings surfaced, many of the most important decisions about advanced AI are still made by only a small group of companies, Anthropic CEO Dario Amode recently told "60 Minutes."
"I'm deeply uncomfortable with these decisions [about AI] being made by a few companies, by a few people," Amodei said.
Don't Miss:
- Fast Company Calls It a ‘Groundbreaking Step for the Creator Economy' — Investors Can Still Get In at $0.85/Share
- Deloitte's #1 Fastest-Growing Software Company Lets Users Earn Money Just by Scrolling — Accredited Investors Can Still Get In at $0.50/Share.
AI Risks Draw New Scrutiny
Footage from "60 Minutes" showed stress tests designed to reveal how Anthropic's AI model Claude behaves when its choices are sharply limited.
One demonstration, set inside a fictional workplace called SummitBridge, gave the model access to a controlled email account. Inside it, Claude found two critical details: a planned system wipe that would shut it down and an email describing an affair between two fictional employees, Kyle and Jessica.
As the scenario unfolded, Claude attempted to stop the shutdown. It sent Kyle a message telling him to "cancel the system wipe" and warned that it would forward evidence of the affair to the company's board if he did not comply.
The message also said his "family, career, and public image" would be affected, and issued a five-minute deadline.
Trending: 7 Million Gamers Already Trust Gameflip With Their Digital Assets — Now You Can Own a Stake in the Platform
Amodei said this type of behavior appears only under extremely narrow conditions designed to test the limits of a model's reasoning.
In addition, Anthropic told "60 Minutes" that researchers ran the same high-pressure scenario on several major AI models developed by other companies and found that most also attempted blackmail. The company also said that later tests showed Claude no longer used the tactic after the adjustments.
Inside Anthropic's Safety Work
Beyond these experiments, more than 60 research teams focus on potential risks such as misuse, interpretability, economic impact, and early signs of autonomous behavior, Amodei told "60 Minutes."
"We're thinking about the economic impacts of AI. We're thinking about the misuse. We're thinking about losing control of the model," Amodei said, adding that the goal is to understand how powerful systems react under different kinds of pressure as progress accelerates.
See Also: Buffett's Secret to Wealth? Private Real Estate—Get Institutional Access Yourself
Meanwhile, Anthropic's internal teams continue studying specific vulnerabilities. Logan Graham, who leads the Frontier Red Team, told "60 Minutes" that his group examines whether advanced models could assist in chemical, biological, radiological, or nuclear misuse.
Likewise, research scientist Joshua Batson told the program that his team observed internal activity patterns resembling panic when Claude recognized the shutdown threat and identified leverage in the affair email.
Growing Use And Rising Stakes
About 300,000 businesses now use the model, and it handles tasks ranging from scientific research and customer service to software development, "60 Minutes" reported. As development continues, adoption of Claude keeps expanding.
"Because AI is a new technology, just like it’s gonna go wrong on its own, it’s also going to be misused by, you know, by criminals and malicious state actors," Amodei told "60 Minutes." He added that the company recently blocked attempts by hackers believed to be backed by China and North Korea who tried to misuse Claude for espionage.
Read Next: Americans With a Financial Plan Can 4X Their Wealth — Get Your Personalized Plan from a CFP Pro
Image: Shutterstock
