In the most famous case to date, a finance employee at engineering firm Arup joined a video call with the CFO and several colleagues (all of them deepfakes) and wired out roughly $25 million across fifteen transfers. The technology that pulled that off is now cheap, fast, and pointed at businesses a fraction of that size.
How little the attacker needs
Three seconds of audio, a voicemail greeting, a conference talk, a social clip, trains a usable voice clone. Your executives' voices are already public. The attack that follows is not sophisticated: a phone call to accounts payable that sounds exactly like the owner, urgent and plausible: "I'm boarding a flight, the acquisition closes today, wire the deposit now, keep it quiet."
Smaller businesses are the growth market, not the exception. A $40K fraudulent transfer from a 30-person contractor is easier to execute and rarely makes the news: the playbook is identical.
Why detection is a losing strategy
"Listen for robotic artifacts" was 2023 advice. Current voice synthesis passes casual inspection, and video is close behind. Any defense that depends on a stressed employee out-detecting a synthetic voice in real time will eventually fail.
The controls that work are procedural, because procedure does not care how good the fake is. One: a payment-verification protocol, meaning every new or changed payment instruction gets a callback to a known-good number, no exceptions, including "the CEO" personally. Two: a drill, simulating a voice-clone request against your own finance team before a criminal does. The first drill almost always finds someone who would have wired the money.
The takeaway
You cannot train ears to beat synthetic voices, but a callback protocol beats them every time. Helm Aware runs the deepfake drill and installs the protocol; it is a half-day that removes your single largest fraud exposure.