Problem
Cyberespionage, defined by the US Cybersecurity and Infrastructure Security Agency (CISA) as “a type of cyberattack in which an unauthorized user attempts to access sensitive or classified data or intellectual property (IP) for economic gain, competitive advantage, or political reasons” [1] is a growing threat for both the public and private sectors from businesses to government agencies. The purpose of cyberespionage is typically for intelligence gathering rather than being conducted for financial reasons [2]. In 2023, more than 1,200 cyberattacks were conducted with the ultimate goal of stealing information [3]. In 2025 alone, Chinese cyber operations skyrocketed by 150% with 11% of cyberattacks globally attributed to cyber espionage [4]. While cyberattacks are often portrayed as highly technically sophisticated, analysis by the European Union Agency for Cybersecurity (ENISA) discovered that 60% of cyberattacks utilized social engineering, such as phishing attacks, to gain initial access into networks [5]. This finding is important because it highlights how human judgment is the most commonly exploited weak point in cybersecurity. As artificial intelligence advances at an unprecedented rate [6], it opens the door to new vectors for social engineering-based cyberespionage. Technologies that have the potential to enhance cyberespionage operations against organizations and agencies are real-time deepfakes (RTDFs) including facial, full-body, and auditory deepfakes.
The United States Government Accountability Office (GAO) defines the term deepfake as “a video, photo, or audio recording that seems real but has been manipulated with AI. The underlying technology can replace faces, manipulate facial expressions, synthesize faces, and synthesize speech” [7]. Cyber threat actors can use this technology to mimic the appearance and voice of ordinary people, celebrities, and political figures. In fact, this has been done in the past to spread disinformation. For instance, a deepfake video of Ukrainian President Volodymyr Zelenskyy announcing Ukraine’s surrender was uploaded to a Ukrainian news site by hackers and proliferated across social media in 2022 [8]. Since then, deepfakes have only become more advanced and more convincing [9]. RTDFs are deepfakes that can be applied during real-time video interactions [10]. Even more sophisticated are full-body deepfakes in which threat actors can digitally superimpose the image of others onto their own to appear as if they were their target [11]. Finally, audio deepfakes can mimic voices in real time [12]. Combining all of these technologies can create a convincing performance, which may be used to fool victims into divulging classified information or credentials to threat actors virtually disguised as superiors or colleagues. The same social engineering exploit that makes phishing so attractive to cyber threat actors is replicated by rapidly developing deepfake technology.
Case Studies
As far back as 2022, the FBI released a public service announcement warning about reports of synthetic identity theft for RTDFs and voice spoofing to interview for remote positions. Many of the positions in question were sensitive, including access to customer personally identifiable information, financial data, corporate IT databases, and proprietary information [16]. AI-enabled synthetic identity spoofing has even been used by North Korean threat actors. For example, an AI image editing platform called Cutout.pro was breached, exposing dozens of suspected North Korean-linked emails using the service to face swap photos [13]. A 2023 investigation by Palo Alto Networks, a U.S.-based multinational cybersecurity company, revealed that DPRK-sponsored threat actors were using RTDFs during interviews, posing as candidates, and attempting to trick interviewers into downloading malware [14]. This tactic is particularly useful to threat actors because the same agent can use it multiple times, with different faces, without revealing their identity. The tool also requires little expertise and basic hardware requirements [15]. Threat actors also utilize synthetic voice mimicry to convince victims to supply confidential information. The FBI documented attackers using this technology to audibly impersonate senior US officials in voice messages, a technique known as vishing. Once rapport was established, the attackers sent malicious links under false pretenses to gain access to their targets’ accounts [17].
Conclusion
While mitigation strategies do exist for RTDFs and vishing attacks, they must constantly adapt to evolving technologies. Moore’s Law stipulates that, as already demonstrated, these technologies will only become more advanced and less expensive. Because social engineering is the linchpin of RTDF and vishing attacks, they are not so easily mitigated by commensurate cyber defense strategies. It is human error that enables this technique, so training of personnel to properly identify suspicious interactions paired with integrated technologies to identify RTDFs and vishing is paramount. Some of the best practices used today are already being made obsolete by rapid development toward deepfake realism and indistinguishability between real and altered faces, bodies, and voices. For example, in 2025, the FBI emphasized that audiovisual inconsistencies, such as hand and feet irregularities, facial distortion, lip-sync errors, and inaccurate shadows or unnatural movements, are indicators of deepfake usage. However, the bureau also noted that the absence of these signs was not necessarily indicative of authenticity because RTDFs and voice spoofers are advancing to the point that they can be indistinguishable from reality [18].
RTDF detection is exceedingly difficult under realistic video-call conditions. A 2025 study testing fourteen detectors found that the highest performing detector’s Area Under the Receiver Operating Characteristic Curve (AUC) declined from 0.804 on high-quality video to 0.560 on low-resolution and heavily compressed videos [19]. AUC, in this case, denotes how well a detector can differentiate a deepfake from genuine real-time video. A score of 0.56 is close to random chance-level discrimination. Thus, by manipulating RTDF video conditions, threat actors may still be able to bypass the most advanced detectors today. However, the ensuing arms race of cyberespionage via RTDFs and detection software will very likely continue to pose a threat to critical confidential information by threat actors unless detectors advance beyond adversarial techniques or legislation regulates RTDF technology, making them more difficult to present as authentic.