Here’s what’s on our plate today:

• 🧪 The AI scam problem has arrived before the defenses.

• 📰 Alexa+ won't stop singing, Lucid's output hits a low, and Amodei's $18M pay.

• 🛠️ Three things worth trying: a verification passphrase, Pindrop, ElevenLabs voice tools.

• 🗳️ Poll: what actually stops a voice-clone fraud?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• The theft: in February, fraudsters took €95M from Fideuram, Intesa Sanpaolo's wealth arm, using a WhatsApp message in the name of Intesa's chief executive and a cloned lawyer's voice to confirm it.

• The fake check: the call looked like a second, independent source, when both came from the same fraudsters.

• The response: Fideuram recovered about €53M, while €36M went through overseas accounts into cryptocurrency.

• The lag: Europol warned about cloned executive voices in 2022, yet refund rules still protect only consumers, and listeners cannot reliably spot a clone.

• The stakes: FBI data cannot yet show AI driving wider fraud losses, but one convincing clone was enough to beat the check the chairman relied on.

The AI scam problem has arrived before the defenses

In The Sting, a 1973 film set in Depression-era Chicago, two con men played by Paul Newman and Robert Redford set out to fleece a racketeer named Doyle Lonnegan. Their scheme depends on a friend at Western Union, the company that transmitted horse-racing results to betting parlors, who can supposedly deliver the winners before bookmakers hear them. Lonnegan is wary enough to demand proof that the friend really works there before placing his bet. The con men oblige by briefly taking over an actual Western Union office and installing one of their crew in the manager's chair. Lonnegan visits, sees a functioning telegraph office, and meets a man behind the right desk. Everything checks out, or so it seems. What he fails to realize is that the entire scene has been staged by the very people he is supposed to be investigating. For most of fraud's history, pulling off such a deception required an elaborate and expensive physical setup.

Executive fraud follows the same principle, but generative AI has made the staged verification much cheaper. In a typical scheme, a criminal impersonates a company's boss and asks an employee to approve a large payment. The usual safeguard is to verify the request with a second person the employee trusts. The logic is straightforward: a fraudster might fake the boss's message, but getting another trusted person to independently confirm it is much harder. AI has changed that equation by allowing the fraudster to impersonate the second person as well, making a familiar voice less reliable as proof of identity. Corporate payments have long relied on employees recognizing the people who give them instructions, so the scam in its new form has arrived before the defenses built to stop it.

That gap became a real bill in February, when fraudsters took €95M from Fideuram, the arm of Italy's largest lender, Intesa Sanpaolo, that manages money for wealthy clients. Reuters reported the theft on September 25, 2026, citing two people familiar with the matter, and both banks declined to comment.

The fraud staged its own confirmation

The scheme began with a WhatsApp message, apparently from Intesa's chief executive, Carlo Messina, asking for urgent help with an overseas transaction. A phone call followed in what Reuters' sources describe as the AI-cloned voice of a senior lawyer, the Italian managing partner of A&O Shearman, who was not involved, according to Cybernews. Fideuram's chairman, Paolo Molesini, then told the finance department to send the money, mainly to accounts in China and Hong Kong.

Each step looked like a precaution being followed, since the message named the most senior person in the group and the call supplied a second trusted voice that agreed with it. To the chairman, that looked like two independent sources, when both came from the same people. Like Lonnegan's visit to the telegraph office, the check ran through a system the fraudsters had built.

The FBI calls this kind of con ‘business email compromise’, and its definition covers phone calls and video meetings as well as email. FinCEN, the U.S. Treasury's financial-crime unit, warned banks in its 2024 alert on deepfake fraud that criminals were using generative AI, including cloned audio and video, to get past the checks banks use to confirm who they are dealing with. The AI part of the Fideuram scheme was a ‘voice clone’, a fake of one real person's voice built from recordings of them. A peer-reviewed study in PLOS One found in September 2025 that listeners struggled to tell clones from real speakers, and they rated some synthetic voices as more trustworthy. For someone taking the call, the natural test of whether a voice sounds right now proves little about who is speaking.

Fideuram could only chase the money after it left

Fideuram's response began only after the money had left the bank, and from that point the task was to get it back. The bank spotted irregularities quickly and alerted banks and authorities in several countries. About €53M was later recovered with help from authorities in China, Portugal, and Italy. The other €36M passed through overseas accounts and was converted into cryptocurrency, and it is still missing. Milan prosecutors are investigating a foreign national living outside Europe, and no Fideuram executive is under investigation. Molesini, the chairman who gave the order, resigned in March, citing personal reasons.

The warnings arrived years before the defenses

Fideuram's loss reaches beyond one bank, because the danger was described well before February. Europol, the EU's police agency, described in a 2022 report on deepfakes a case in which criminals used a cloned executive's voice to make an employee transfer $35M. It advised organizations that authorize by voice or face to strengthen the whole authorization process over relying on deepfake detection alone. The FBI warned in December 2024 that generative AI lets criminals work at a larger scale with less time and effort. By then, Arup, the British engineering firm, had lost about $25M after a video call in which a Hong Kong employee faced a deepfaked finance chief.

The protections built since then mostly guard consumers, because refund rules in the UK and the EU's agreed payments rules put more of the loss on banks when customers are tricked into paying. Those rules do not reach a chairman moving his own institution's money, so a company fooled at the top carries the loss itself. Listening harder is no reliable defense either, as the PLOS One study shows, which leaves the approval process as the place where the defense has to change.

The one defense in these cases that stopped a clone came from an executive at Ferrari, the Italian carmaker, in July 2024. When a caller using a cloned voice of chief executive Benedetto Vigna reached him, he said, “Sorry, Benedetto, but I need to identify you.” He then asked which book Vigna had just recommended, and the call ended. The question rested on something only the real Vigna knew, and no recording of his voice could supply the answer, which is the shift Europol had described two years earlier.

The loss data has not caught up with the cases

If the AI scam problem has truly arrived, the loss figures should show it, and so far they barely do. The FBI measures this kind of fraud through complaints from victims. Reported business email compromise losses stayed close to $3B a year from 2023 to 2025. Complaints that mentioned AI made up about $30M of the 2025 total, roughly 1%.

On that evidence, Fideuram looks like an old-fashioned scam built on rank. The request carried the name of the parent company's chief executive, and that name did most of the work. The cloned voice only added polish.

The weakness in the figures is how they are collected. The FBI counts a complaint as AI-related only when the victim notices AI and reports it, and it started counting these cases only in 2025. The PLOS One study found that listeners struggle to tell a cloned voice from a real one, so many victims of AI scams would report an ordinary fraud. The figures therefore show how many AI fraud victims can spot, which may be much less than how many exist. They cannot yet settle how much fraud AI is causing. They also leave the Fideuram case unchanged, since one convincing clone was enough to beat the check the chairman relied on.

The staged check no longer needs a crew

Lonnegan trusted a check he could see, in an office the con men had borrowed and staffed. Fideuram's chairman trusted a check he could hear, in the voice of a lawyer who knew nothing about the call. The office needed a crew and a building, while the voice needed only recordings of the right person, and the warnings about that shift have been public since 2022. With the staged check now that cheap, how many companies will rebuild their approvals before the next call arrives, and how many will wait for their own €95M lesson?

Quick Bits, No Fluff

• Alexa+ won't stop singing: Amazon's new Alexa+ has been creepily singing "lalala" for minutes on end, an unsettling glitch that's denting trust in its upgraded assistant.

• Lucid's output hits a low: Lucid Motors' EV production fell to its lowest level in nearly two years, deepening doubts about the luxury automaker's path to scale.

• Amodei's $18M pay: Anthropic's Dario Amodei made $18M last year, landing mid-pack among tech CEOs, a modest figure for one of AI's most prominent leaders.

3 Things Worth Trying

• Set a verification passphrase: Agree on a shared code word with family or your finance team — the Ferrari trick that beat a voice clone — so a familiar voice alone can't authorize anything.

• Pindrop: Voice-authentication and deepfake-detection tech built for call centers, a useful look at how enterprises are trying to verify who's actually on the line.

• ElevenLabs voice tools: Try cloning your own voice to see firsthand how little audio it takes; the fastest way to understand why "it sounded like them" is no longer proof.

Thursday Poll

The Toolkit

• Leonardo AI: AI image and video generator with fine-grained creative controls, built for designers, marketers, and game studios who need consistent style at scale.

• Modal: Serverless cloud for running Python and AI workloads; lets you spin up GPUs in seconds without touching infrastructure.

• Quillbot: AI writing assistant that paraphrases, summarizes, and rewrites text on demand, useful for tightening drafts or escaping your own voice.

Rate This Edition