A WhatsApp message from the boss. A follow-up call from a lawyer whose voice you know. An urgent, confidential transfer. That's all it took to move about €95 million out of one of Italy's biggest private banks. If a bank chairman fell for it, your shop's WhatsApp is not too small to be next.
Today's digest is mostly about trust: who you trust in a chat, whose search results send you customers, and how much access you hand to AI agents. Four stories, then quick takes.
A fake WhatsApp and a cloned voice drained €95 million from a bank
Here's the one you can't ignore. Fideuram, the private-banking arm of Italy's Intesa Sanpaolo, lost about €95 million to a scam that started with a WhatsApp message that looked like it came from the group's CEO. He asked the bank's then-chairman for help with an urgent overseas deal. Then came a phone call from a "senior law-firm partner." The voice was cloned with AI.
The scheme ran in February 2026 and only came out publicly in the last few days. Investigators in Italy, China and Portugal got back roughly €53 million. At least €36 million was turned into crypto and is still being traced. The chairman resigned in March; prosecutors in Milan are investigating a suspect living outside Europe.
Why it matters for you: scammers don't need a bank to make money. The same trick works on a small shop: a "supplier" asking for payment to a new card number, a "courier" sending a voice note, a "customer" who wants a refund to a different account. My rules, starting today:
- No money moves on a chat message alone. Call back on a number you already had, not the one in the message.
- A familiar voice is no longer proof. A few seconds of audio from your stories is enough to clone it.
- Urgency plus secrecy is the red flag. Real partners can wait ten minutes for you to check.
Also tell your staff. The person answering your WhatsApp business inbox is the one who'll get the fake "it's me, send it now" message, not you.
Google's AI Mode sends fewer people to websites, and the court won't stop it
Two pieces of news landed together and they tell one story. First, researchers from the University of Pennsylvania and Northeastern ran a randomized experiment with about 1,100 real Google users in their everyday browsing. When people were put on Google's AI Mode, their click-through rate to outside websites fell by 18.8 percentage points. They also searched less, rated Search as less useful, and were more likely to try Bing or DuckDuckGo.
Second, on October 1 a US federal judge, Amit Mehta, dismissed the antitrust cases that Penske Media and Chegg brought against Google over AI Overviews. He accepted that publishers are losing traffic, but said antitrust law can't replace lawmakers here. His line: "an expectation is not an agreement."
Why it matters for you: if your shop site lives on Google traffic, nobody is coming to rescue those clicks. Fewer people will land on your pages, so each visitor counts more. The sellers who win will turn that one visit into a conversation, with a website chat that actually sells, and will build channels Google doesn't control: your Instagram page, your WhatsApp list, your Telegram channel.
Apple puts a fence around what AI agents can read on a Mac
Apple says it's tightening macOS "Full Disk Access", the permission that lets an app read everything: files, mail, messages, browser history. It was built for backup apps. Apple now says AI agents make that level of access much riskier, and that some developers are using it in ways that expose users "without users' full knowledge."
From now on, an app that wants full access will need a fresh, explicit action from the user, not a permission you clicked once and forgot. It came days after a journalist said Meta's Muse app on Mac had read their private messages, a claim Meta disputes. That's the same Muse that gave a seller's address to a buyer earlier this week.
Why it matters for you: your laptop holds customer chats, addresses and supplier invoices. Before you give any AI app "access to everything," ask what it actually needs. A good assistant works with the data you choose to share, not your whole disk. Apple making this harder is a favor to you.
Amazon turns its ad console into a chatbot
Amazon merged its ads console and its DSP into one platform called the Amazon Ads Agent. You talk to it in plain language, it remembers context across campaigns, and it can plan, build and optimize ads across search, streaming TV, video, display and audio. A new option, DVA+, hands over targeting, bidding and even creative to the AI.
Amazon says it built this for smaller advertisers who don't have a media team. Reporting works in plain questions too, no SQL.
Why it matters for you: even if you never sell on Amazon, this is where ad tools are going: you state the goal, the agent runs the campaign. That makes good ads cheaper to run, which means your competitors will run more of them. The edge moves to what happens after the click. If the ad works and nobody answers the DM, you paid for nothing.
Quick takes
- Microsoft: attackers are winning the early AI race. Its 2026 Digital Defense Report says the median time from a flaw being found to being exploited has dropped below 24 hours, and the first fully autonomous ransomware hit real organizations in July. Update your apps and phone when asked. Really.
- Inworld bought Ultravox, a platform for real-time voice agents. Inworld's voice model supports 200+ languages with very low lag. Voice agents that answer calls naturally are getting closer, and cheaper.
- A light-based chip from UCLA spots deepfake video with about 98% accuracy on a standard test set, across 15+ streams at once. It's research, not a product, but detection is finally catching up to the fakes in our lead story.
- Another safety exit at OpenAI. David Robinson, who led its safety transparency work, has left, days after three safety researchers were fired. Worth watching, not panicking over.
The thread through today: trust is now the scarce thing. Customers can't trust a voice, Google won't send you as many visitors, and every agent wants more access. The shops that do well will answer every real customer fast, in their own channels, without handing their data to the wrong thing. That's the job Vardast does: it answers your Instagram, WhatsApp and Telegram customers around the clock and follows the rules you set.