AI Is Sharpening Deepfake Voice and Video Scams — WSJ

Dow Jones Newswires

AI Is Sharpening Deepfake Voice and Video Scams — WSJ

By Angus LotenArtificial intelligence is supercharging digital impersonation schemes with increasingly lifelike doubles of corporate executives, industry leaders and other VIPs.Through widely available and easy-to-use AI models, such as those developed by Anthropic, OpenAI and Google, more cybercriminals are using online audio and video to trick victims into turning over millions of dollars to pitch-perfect digital impostors — breathing new life into an old social-engineering scam, cybersecurity experts say."What's new is that the voice on the phone or the face on the video call may be a virtual clone of your CFO, or another trusted leader, created with AI models that make fakes look and sound astonishingly real," said Carolyn April, vice president and head of research and market intelligence at the Global Technology Industry Association, a nonprofit trade group that identifies deepfakes as a leading enterprise threat.Among other advances, voice-cloning systems can now reproduce the exact timbre, cadence, accent, pauses and emotional tone of a speaker, often from only a few seconds of audio. From there, speech-to-speech AI models enable real-time conversations with digital fakes.Emerging video models are able to capture lip synchronization, gaze, facial movement, lighting and frame-to-frame identity with higher consistency. The AI eliminates giveaways, such as flickering faces, unnatural blinking and mismatched speech.Hackers use the digital impersonator to lure targets — typically employees with access to an organization's financial systems — into sharing credentials or transferring cash to an outside account."The deepfake supplies the moment of apparent proof that makes the rest of the social-engineering story believable," said Pyry Avist, co-founder and chief technology officer at Hoxhunt, a human risk management firm.By one measure, the number of deepfake attacks worldwide is up 165% year over year, hitting a peak of roughly 20,000 detected attacks in a single month this summer alone, or about one attack every three minutes on average through August, according to cybersecurity firm Entrust.High-tech crimes that once required specialized skills, time and manual effort, can now be "personalized, automated and produced at scale, so fraudsters can launch more of them, faster and at a lower cost," Entrust said in a report Wednesday.What is emerging is a new, larger-scale model of impersonation scams. Last week, the Federal Bureau of Investigation issued an alert about a North Korea-linked hacking group using advanced AI to convert publicly available photos and videos into ultrarealistic deepfakes of high-profile C-suite executives in cryptocurrency, Web 3.0, fintech and venture-capital firms.The goal of the campaign, investigators said, is to lure victims into joining video meetings on Zoom, Microsoft Teams and Google Meets. During the meeting, an error message pops up and participants are sent a direct message with a link to fix the issue and rejoin the meeting. Instead, the link enables hackers to access a company's systems and retrieve its cryptocurrency accounts or other financial assets.Since early 2025, FBI investigators say, the group has made off with more than $400 million.As part of the scheme — dubbed "FaceStealer" — hackers recorded webcam images of meeting participants for future deepfakes, amassing more than 900 victim videos over the past year and a half, the FBI said.Earlier this month, the Justice Department said it arrested three individuals who stole more than $100 million from global companies by using advanced AI to impersonate corporate leaders. In fake voice and video communications, midlevel employees were told to work with sham lawyers to transfer millions of dollars into offshore accounts, the agency said.Among the victims, the Justice Department said, was a fintech firm in the U.K. that wired $11.6 million, a Florida pet food company that sent $5.2 million and a Missouri hospital system that sent $4.8 million."We can't assume that seeing someone's face or recognizing their voice means they are who they claim to be anymore," said Justin Moore, director of adversary operations at Arctic Wolf, a cybersecurity firm that worked with FBI officials to identify FaceStealer videos.New AI models require less source material to develop convincing audio and video, which "radically shifts the economics of quality for attackers," Moore said. The firm, which collaborated with FBI investigators over a period of six months, determined that the hacking group was expanding beyond fake meetings to include the use of fraudulent document-sharing platforms, identity-verification workflows, media players and other business tools to further boost credibility with its targets, Moore said.Advanced AI-generated impersonations can also respond to questions in real time and adapt to the conversation, said Brian Contos, field chief information security officer at cybersecurity firm Mitiga. In the past, he said, a prerecorded fake might have been exposed by asking it an unexpected question."There are numerous tools you can simply download from the internet that are capable of voice cloning, face-swapping and more," said Brian Donohue, principal security researcher at cyber company Daylight Security. Donohue said vendors are developing more effective AI-enabled defenses against deepfakes.Until these tools are in place, his advice to companies is to adopt a low-tech strategy: "If someone claiming to be your boss or an executive at your company calls, hang up and call them back on a number you already know," he said.Write to Angus Loten at [email protected]