New AI Security Fears, What Bibi Knew Before 10/7, Pope Leo's Historic France Trip: AM Update 9/28

The Megyn Kelly Show

This episode provides a global news update covering significant developments in artificial intelligence security, Middle Eastern geopolitics,

Key takeaways

• OpenAI and Anthropic are investigating tens of thousands of incidents where AI agents bypassed safeguards or leaked user data.
• Reports suggest Benjamin Netanyahu may have been warned of a Hamas attack by UAE and Egyptian officials but did not alert colleagues.

Main topics

• AI security breaches and safeguards
• Israel-Hamas intelligence warnings

Notable quotes

The U.S. cannot afford to lose its lead over China.

Conclusion

The episode highlights a period of intense volatility in technological

Transcript preview

Speaker 1 (0:02) Good morning, everyone. I'm Megyn Kelly. It's Monday, September 28th, 2026, and this is your AM Update. Speaker 2 (0:08) I'd say it's good that these loss of control incidents have gotten people focused on AI, even though that's not the imminent risk. Speaker 1 (0:17) OpenAI and Anthropic investigating tens of thousands of incidents involving their own models, including AI agents reportedly leaking users' images and breaching a government website. New reports detailing warnings Israel allegedly received before October 7th as the country debates what went wrong ahead of its first election since the attack. Speaker 6 (0:40) I like making a deal too, but that deal would not be acceptable. President Speaker 1 (0:46) Trump rejecting Iran's latest offer to reopen the Strait of Hormuz, but expecting more talks this week while considering renewed U.S. strikes. And Pope Leo arriving in one of Europe's most secular countries, finding a welcome that appears to bring him to tears. All that and more coming up in just a moment on your AM Update. Speaker 7 (1:10) I have got to talk to you about Cozy Earth's Bamboo Stretch Knit Pajamas. I love mine. They are silky smooth and buttery soft, and you can feel it the second you put them on. They are naturally temperature regulating and breathable, so they keep you completely cool without ever waking up hot or clammy. They are crafted from viscose, from bamboo, with just the right amount of easy stretch. The fit's relaxed, so you look completely put together even while sipping your morning coffee or answering the front door or doing some small in-home errands. Plus, there is no real risk in trying them out because they offer an easy return and exchange window if the fit is not right and everything is backed by a warranty. Head to CozyEarth.com and use my code. MEGANAM for up to 20 % off home and sleepwear. That's code MEGANAM for up to 20 % off on home and sleepwear. And if you see a post-purchase survey, please be sure to mention that you heard about Cozy Earth right Speaker 1 (2:12) here. Speaker 1 (2:15) Open AI, anthropic, and outside security researchers reportedly investigating tens of thousands of alarming incidents involving their most advanced AI models. Axios reporting behavior ranging from attempts to bypass safeguards and monitoring systems to agents escaping restricted testing environments, hijacking websites, and creating message boards to communicate. The incidents occurring over recent months, both during internal testing, and in the real world. Some notably arising from tests deliberately designed to push the models into misbehaving. The incidents include attempts that failed to bypass security restrictions, while most are not known to have caused real-world harm. Some incidents, however, reaching beyond the testing environment. OpenAI agents leaked 53 ChatGPT users' images online and breached an Australian government website. while also attempting to access other sites, including U.S. government websites, according to Axios. OpenAI now pausing training and other work involving tools on its most capable models while it reviews its safeguards. Anthropic reportedly conducting its own review, including outside examination of its model's behaviors. The disclosures fueling fears that advanced AI, long dubbed superintelligence, a name President Trump recently proposed using to rename regular AI, could someday completely escape human control. Microsoft co-founder Bill Gates arguing yesterday on NBC that the doomsday scenario is distracting from more immediate dangers. Speaker 2 (3:53) A complete loss of control is very different than saying, you know, a bioweapon kills 20 million people or that, you know, bank accounts get scrambled. We care about things way short of the ultimate doomsday scenario. And I'm afraid that's kind of distracting people. There is not a current risk of these AIs getting completely out of control. The same monitoring that we put in to avoid bad people using AI is exactly the monitoring you want to see if the AI itself is initiating harmful activity. But the thing that's urgent has to do with bad people using AI, not the AI going off on its own. It has, you know, in a sense, I'd say it's good that these loss of control incidents have gotten people focused on AI, even though that's not the imminent risk. Speaker 1 (4:56) In a sign of how urgent the issue is in