In January 2024, an employee at the engineering firm Arup joined a video call with people he believed were his colleagues, including the company’s chief financial officer. Everyone on the call was fake. He authorized transfers totaling roughly $25 million before anyone noticed, and the Financial Times later confirmed the scale of the loss.
The Barrier to Entry Dropped
Cloning voices formerly required a studio recording with a willing participant. Today cloning voices requires nothing but an audio clip from a podcast.
Synthetic media costs have fallen along with its technological barriers. Trend Micro estimated in April 2026 that an attacker could assemble a polished scam operation using inexpensive AI and automation tools for as little as $60 per month. The company also reported that some criminal services bundle voice cloning, deepfake video, website generation, and targeted messaging into ready-made kits. Thus, even if an attacker is successful a small percentage of the time, such attacks may remain financially viable.
In December 2024, the FBI released a public service announcement indicating that the cost reduction associated with using generative tools has reduced the barrier to entry particularly for impersonation fraud. There are many examples of social media providing sufficient training data for the various forms of synthetic media. These include earnings calls, panel discussions, birthday videos, and voicemail greetings. No one intended for these items to become biometric data; yet today they are.
Where Do the Fraudulent Attempts Actually Hit?
They do not hit the flashy targets. Instead, most fraudulent activity hits contact centers and help desks; this is due to the fact that this is typically where there is a person who has the capability to reset your password, turn off MFA or allow you to complete a stuck transaction.
According to Pindrop’s (2025) study, there was a significant increase in deepfake-related fraud activities directed towards call centers. Deepfake fraud activities against call centers increased over 1300 percent throughout 2024. Pindrop also noted significant increases in deepfake-related fraud activities at both banks and insurance companies.
Agents are ill-equipped to deal with this type of situation and it is unreasonable to assume they would be.
Why Payouts Are Re-Verified
Outgoing transfers are scrutinized much more than incoming deposits.
The response across regulated finance has been to stop treating identity as something you prove once. Guidance on how to withdraw money from Bovada shows a platform-specific process: withdrawal requests are reviewed, documentation may be requested, and available methods may depend on how the account was funded. Cryptocurrency withdrawals require a new wallet address for each request. A voice on the phone cannot override that binding. Neither can a convincing face on a video call, because the check is not asking who you appear to be. It is asking whether this destination was registered before the request arrived.
This method was in place before deep fakes existed.
How Much Does Detection Really Help?
In lab tests, vendors claim extremely high success rates. In lab test environments these claims are true. However, when dealing with live scenarios the environment will always be dirty. Therefore, even a slight degradation of signal (i.e. compression, background noise, packet loss), can cause the performance of detectors to plummet.
Additionally, there exists a significant arms race problem. Current detection methods are trained using current generations’ techniques. When a new generation method appears, existing detectors may initially perform less reliably against it. Once again, the “gap” opens. Gartner conducted a survey of cyber security professionals in September 2025 and concluded that over half of respondents indicated they had experienced at least one form of deepfake related attack within the previous 12 months. Additionally, Gartner indicated that audio calls were the most prevalent vector of attack.
Likewise, the money follows a similar curve. Deloitte’s financial services research indicates that U.S.-based generative-AI enabled fraud is expected to climb from approximately $12.3 billion in 2023 to almost $40 billion by 2027. While projections should be viewed skeptically, the trend is clear.
Thus, the industry has shifted away from relying solely on detection for protection. Instead, the focus has shifted towards developing controls that detect or prevent a fraudulent transaction without necessarily detecting whether a call is genuine or a fake.
Out-of-Band Verification and Dual Approval
Practical countermeasures to these types of attacks are far less glamorous and many of them date back decades. These include:
- Out-of-band approval: verify a high-risk request through a communication channel other than how your contact initiated contact.
- Second approver. For requests greater than a certain amount of money, require that two individuals approve the action.
- Callback verification. Hang up and dial a number you already had, not one the caller supplied.
These issues compound with injection-based attacks. An attacker does not simply present a fake face for authentication purposes; instead, the attacker injects the video stream directly into the authentication application thereby circumventing the camera altogether. To combat this type of behavior providers have implemented liveness checks in order to determine if a legitimate camera was utilized during the authentication process. However, this is also contested territory.
None of these alternatives restore the prior assumption that knowing someone’s voice means you know who they are. Unfortunately, what now represents identity proofing is far less exciting: verified destinations, secondary channels and a desire to slow down a transaction that appears to be time-sensitive. Fraud departments have long recognized that urgency is typically indicative of fraud; unfortunately the technology has changed while that fundamental truth remains constant.
