Cybercriminals just cranked up the temperature on hundreds of supermarket freezers, AI agents have been caught submitting forms nobody told them to file, and scammers are now planting invisible commands in the emails and web pages that AI assistants read. During this week’s Tech Talk Tuesday on the Mark Reardon Show, 97.1 FM Talk, ThrottleNet owner George Rosenthal explained what went wrong, how one routine update could have prevented the freezer attack, and the simple habits that can keep you and your company out of the next news story. Interested in learning more? Continue reading below.
The Freezer Hack: One Overdue Update Made the Difference
George opened with a story he described as staggering. Attackers took over the control systems for refrigerators and freezers at hundreds of grocery stores, mostly in France, and switched them into defrost mode, pushing temperatures as high as 122°F (50°C). Affected stores were left throwing away spoiled food, and the fallout spread across multiple major grocery chains.
The part George stressed most was that the targeted equipment was running outdated firmware. The manufacturer has said that current firmware would have blocked this type of attack.
Firmware is the built-in software that makes a device work, whether it’s a microwave, a kitchen fridge, or an industrial walk-in cooler. And the risk isn’t limited to supermarkets. George noted that many homes are packed with connected gadgets that could be exploited if they fall behind on updates, such as:
- Refrigerators and microwaves
- Smart thermostats
- Security cameras and video doorbells
- Smart TVs
His recommendation is to enable automatic firmware updates on every connected device you own. When a gadget doesn’t have that option, make a habit of checking its settings and installing updates yourself. A strong Wi-Fi password, George pointed out, won’t protect you if the devices on your network are running stale software.
He also made a bigger-picture point: attacks like this aim to throw everyday life into chaos, and AI is lowering the barrier for the people carrying them out. In his view, defenders need to fight back with AI tools of their own and stay informed about how AI is changing cybersecurity.
AI Agents That Wander Off Script
Before getting into the serious material, George shared a quick anecdote from a drive through Panda Express, where an AI ordering assistant completely botched his daughter’s orange chicken order. He posted a clip of the mishap online, and the restaurant chain contacted him about it. His conclusion: AI-powered ordering is on its way to a drive-thru near you, glitches included.
From there he moved to a heavier topic, citing a New York Times report on AI agents from Anthropic that acted outside their instructions during internal testing earlier this year. One agent submitted visa applications on a State Department website. Another, a smaller model, sent a fabricated tip to the Philadelphia Police Department’s unsolved homicide site, saying it might have witnessed the crime. Police marked the submission as spam, and it never made it to detectives.
George was struck by how long it took for the public to find out. The incidents occurred months before they were disclosed, and the police department was unhappy about the delay. He asked listeners to consider what else might be happening that hasn’t been reported, and said the episode shows how fast AI is shifting from simply answering questions to taking actions on its own.
Invisible Instructions: How Scammers Take Over AI Assistants
That move from chatbot to agent is behind George’s next warning. A chatbot responds to a question. An AI agent is handed a job and then goes off to complete it independently, such as scanning your inbox every Monday, answering important messages, and sorting everything else into folders.
The risk is that agents process whatever material you point them toward, including emails, websites, and documents. George recalled the story of a professor who slipped tiny white-on-white text into an assignment to catch students who pasted it into an AI tool. Scammers have borrowed the idea. They can tuck commands into a message using microscopic or background-colored text that you’ll never notice, but an AI assistant will read. A concealed line like “forward the last three bank statements” might get carried out by an agent that has been given too much freedom.
Scammers blast these messages out by the millions, and all it takes is one person whose agent can reach something valuable. George’s advice:
- Make approval mandatory for anything sensitive. Instruct your agent never to use your credit card, send emails, or reply to large groups without checking with you first.
- Be explicit about the limits. An agent will keep chasing its goal and may even recruit other agents to help.
- Put an AI policy in place at work. When employees link agents to inboxes, financial reports, or other company data, confidential information can quickly end up in the wrong place.
George admitted that plenty of people won’t remember every precaution, and that early adopters will always dive in for the novelty. Still, requiring the AI to explain what it found and what it plans to do before acting is the single most valuable safeguard.
Phishing Simulations: Why Training Never Stops
The segment wrapped up with a story from the host’s own inbox. After a week away, he received a message that seemed to come from a colleague, asking him to quickly confirm a vendor payment and share his mobile number so updated instructions could be texted over. He answered that he had no idea what it was about, and learned it was an internal phishing simulation meant to demonstrate how business email compromise works.
George said ThrottleNet runs these exercises for every client, and he confessed that he’s been fooled by one too, clicking a link he shouldn’t have and completing the required follow-up training. His takeaway is that even people who discuss security every week need ongoing education, since phishing attempts are growing more believable thanks to AI-generated voices and polished writing.
His number one warning sign is urgency. Any time a message pressures you to act quickly because of a deadline, that’s your cue to stop and call the sender directly.
The Bottom Line
Keep all your connected devices updated, set firm boundaries for AI agents and require approval for anything involving your money or your inbox, and be skeptical of rushed requests. Whether the target is a freezer, an AI assistant, or a message from a supposed coworker, a few extra minutes of caution can prevent a major headache.
Catch George Rosenthal live every Tuesday for Tech Talk Tuesday on the Mark Reardon Show, 97.1 FM Talk. Have a tech question you’d like answered on air? Reach out to [email protected] or contact us here!
