How deepfakes went from party tricks to national security threats

A mid-level finance worker at Arup, a global engineering giant, sat through a video conference in early 2024 with his Chief Financial Officer and several other colleagues to authorize a secret transaction. Every person on that screen looked right, sounded right, and blinked at the right intervals, yet every single one of them was a synthetic digital puppet. By the time the call ended, the employee had wired $25.6 million into five different bank accounts, executing perhaps the most successful corporate heist in the history of visual deception.

The Hong Kong police later confirmed that the entire "meeting" was a pre-recorded deepfake loop. This wasn't a grainy phishing email or a suspicious link; it was a high-definition assault on the very concept of visual evidence. We have officially exited the era where "seeing is believing" served as a baseline for human interaction. Today, the cost of a sophisticated corporate coup or a national security breach has dropped to the price of a monthly subscription to a high-end GPU cluster.

🎭 The Reddit lab leak

Deepfakes did not begin in a government bunker or a prestigious university lab. They began in late 2017 with a Reddit user named "deepfakes" who decided to superimpose the faces of Gal Gadot and Scarlett Johansson onto the bodies of adult film stars. This wasn't just a crude Photoshop job; it was the first public application of Generative Adversarial Networks (GANs), a machine learning framework designed by Ian Goodfellow in 2014. GANs work by pitting two neural networks against each other: one creates the fake, and the other tries to spot it, forcing the creator to improve until the deception is statistically perfect.

The code used for these early celebrity face-swaps was quickly open-sourced, leading to the creation of FakeApp and later DeepFaceLab. What was once a niche academic pursuit became a plaything for the internet’s most bored and malicious actors. For the first few years, the results were uncanny but flawed—eyes didn't blink correctly, and skin textures looked like wet marble. But the pace of iteration in machine learning is exponential, not linear. By 2021, a TikTok account named @DeepTomCruise, created by VFX artist Chris Ume and Cruise impersonator Miles Fisher, showed the world that with enough data and talent, the uncanny valley could be bridged entirely.

The transition from "party trick" to "existential threat" happened the moment these tools moved from the hands of hobbyists to the toolkits of state-sponsored psychological operations units. In 2022, a low-quality video of Ukrainian President Volodymyr Zelenskyy surfaced on social media, purportedly telling his troops to surrender to Russian forces. While that specific attempt was clumsy and easily debunked, it served as a proof-of-concept for the next decade of warfare. Kinetic strikes are expensive and risky; a well-timed deepfake is cheap, viral, and carries the potential to shatter a nation’s morale before a single shot is fired.

📉 Market manipulation at the speed of fiber

Financial markets are particularly allergic to uncertainty, a fact that deepfake operators are beginning to exploit with surgical precision. In May 2023, an AI-generated image of a massive explosion at the Pentagon began circulating on X, the platform formerly known as Twitter. Despite the image containing several tell-tale AI artifacts—distorted pillars and a fence that melted into the sidewalk—it was shared by an account with a blue checkmark posing as Bloomberg Feed. Within minutes, the S&P 500 dipped by 30 points, temporarily wiping out billions of dollars in market capitalization before the Department of Defense could issue a formal denial.

This incident revealed a critical vulnerability in the global financial infrastructure: the speed of algorithmic trading. Modern hedge funds use "sentiment analysis" bots that scrape social media for keywords and images to make split-second trades. A deepfake doesn't need to fool a human for hours; it only needs to fool a bot for thirty seconds to trigger a massive sell-off. We are looking at a future where "Flash Crashes" aren't caused by technical glitches, but by synthetic media specifically designed to trigger the automated anxieties of the NYSE.

The threat extends into the boardrooms of the Fortune 500. Short-sellers could easily deploy a deepfake of a CEO making a racist remark or confessing to an SEC investigation. By the time the company’s PR department can get the real CEO in front of a live camera to prove the video was a forgery, the stock has already tanked, the short positions have been closed, and the attackers have vanished into the anonymity of the blockchain. The friction of the truth is no match for the velocity of a believable lie.

🗳️ The 48-hour democracy killer

The most dangerous window for a deepfake is the "quiet period" before an election—the final 48 to 72 hours when most voters are making up their minds and media outlets are scrambling to cover every breaking story. In September 2023, just two days before Slovakia’s parliamentary elections, an audio recording appeared on Facebook. It featured Michal Šimečka, the leader of the pro-NATO Progressive Slovakia party, apparently discussing how to rig the election by buying votes from the country’s Roma minority and joked about raising the price of beer.

Šimečka immediately denounced the audio as a fake, and fact-checkers at Agence France-Presse confirmed it was synthetic. But the damage was done. The recording went viral on Telegram and WhatsApp—dark social channels where debunking is nearly impossible. Šimečka’s party narrowly lost the election to Robert Fico, a populist with pro-Russian leanings. This wasn't just a political defeat; it was a demonstration of how a few megabytes of audio can alter the geopolitical trajectory of a NATO member state.

The United States is not immune. In January 2024, thousands of New Hampshire voters received a robocall that sounded exactly like President Joe Biden, telling them to stay home and "save your vote" for the November general election. The call used a voice-cloning tool from ElevenLabs, a startup valued at $1.1 billion that can replicate a human voice with 99% accuracy using only 30 seconds of reference audio. The FCC eventually traced the calls to a political consultant named Steve Kramer, who was hit with a $6 million fine, but the precedent was set. The barrier to entry for voter suppression has been lowered to the cost of a burner phone and a $20-a-month AI subscription.

💻 The industrialization of identity theft

Beyond the high-stakes world of elections and stock markets, deepfakes are being used to dismantle the basic security protocols of the digital economy. Most modern banking apps use "Liveness Detection" or "Video KYC" (Know Your Customer) to verify a user’s identity. You hold your phone up, blink, turn your head, and the app confirms you are a real human. In 2023, researchers at Sensity, a visual threat intelligence firm, found that deepfake tools are now capable of bypassing these "liveness" tests in real-time.

The dark web is currently flooded with "deepfake-as-a-service" providers. For a few hundred dollars, a criminal can buy a custom-made digital mask that can be mapped onto their face during a live video call. This allows them to open fraudulent bank accounts, take out loans in someone else's name, or gain access to secure corporate networks. We are seeing a shift from bulk phishing—where a hacker sends a million emails hoping for one click—to "spear-phishing on steroids," where the hacker calls you on FaceTime looking and sounding exactly like your daughter, claiming she’s been in a car accident and needs an immediate Zelle transfer for the hospital bill.

The psychological toll of this is immeasurable. When we can no longer trust our ears or eyes during a personal communication, the social contract begins to fray. The FBI has already warned about the rise of "virtual kidnappings," where parents are played synthetic audio of their children screaming for help. These aren't just technical exploits; they are predatory attacks on the human nervous system.

🏛️ The Liar’s Dividend

Perhaps the most insidious threat of deepfakes isn't that people will believe things that are false, but that they will stop believing things that are true. Legal scholars Danielle Citron and Robert Chesney call this "The Liar’s Dividend." In a world where deepfakes are known to exist, any public figure caught in a genuine scandal can simply claim the incriminating evidence is an AI-generated forgery.

We saw the first stirrings of this in the 2024 U.S. election cycle. When unflattering videos of candidates surfaced, their supporters immediately labeled them "cheap fakes" or AI manipulations, regardless of their authenticity. This creates a "choose your own reality" environment where objective truth becomes a matter of political tribalism. If a politician is caught on tape taking a bribe, they no longer need to explain the bribe; they only need to sow enough doubt about the video's provenance to satisfy their base.

This reality is a gold mine for authoritarian regimes. In 2023, the Chinese government introduced some of the world's first regulations requiring deepfakes to be clearly labeled, but the primary goal was not to protect the truth. It was to ensure the state maintains a monopoly on the creation of reality. By controlling the "truth-checking" infrastructure, the state can retroactively label any dissident video as a deepfake, effectively erasing real-world protests or human rights abuses from the digital record.

🛡️ The arms race of the authentic

How do we fight back against a ghost? The current strategy is a two-front war: detection and provenance. On the detection side, companies like Reality Defender and Intel are developing "FakeCatcher" technology that analyzes video for biological signals that AI still struggles to replicate, such as the subtle change in skin color caused by blood pumping through veins (photoplethysmography). If a face on a screen doesn't have a pulse, it’s a fake.

However, detection is a reactive game. Every time a detection tool gets better, the GANs used to create deepfakes simply incorporate that feedback to become even more realistic. It is a perpetual motion machine of deception. This has led the industry toward "provenance"—the idea that we should focus on certifying what is real rather than trying to spot what is fake. The C2PA (Coalition for Content Provenance and Authenticity), backed by Adobe, Microsoft, and Sony, is building a "digital nutrition label" for content. Using cryptographic metadata, a photo or video can be tracked from the moment the shutter clicks to the moment it appears on your screen, proving it hasn't been tampered with.

But provenance requires universal adoption, and the most dangerous actors—the troll farms in St. Petersburg and the hackers in North Korea—will never use it. We are entering a fractured media environment where "verified" content exists in a gated garden, while the rest of the internet remains a wild west of synthetic hallucinations. For the average person, the burden of skepticism has never been higher, and the tools for verification have never been more complex.

🔍 The zero-trust future

The endgame of the deepfake era isn't a world of total lies, but a world of zero trust. We are moving toward a "Post-Media" environment where the default assumption for any digital asset is that it is fraudulent until proven otherwise. This will fundamentally change how we conduct business, how we run governments, and how we interact as a species. The high-trust society that allowed the internet to flourish is being replaced by a low-trust survivalist mode where "seeing" is just the beginning of a long and expensive verification process.

National security in the 2020s won't be defined by the size of a country's nuclear arsenal, but by the resilience of its information ecosystem. A nation that cannot agree on what its leaders said yesterday cannot function as a democracy. We are currently losing the battle for reality, not because the technology is too good, but because our institutions are too slow. The $25 million Hong Kong heist was a warning shot; the next one will be aimed at the heart of the global order.

Ultimately, the deepfake threat forces us to return to the only thing that cannot be easily synthesized: physical presence and long-term reputation. In an age of infinite digital replicas, the authentic human becomes the ultimate premium asset. We may find ourselves retreating from the digital town square, back into smaller, verified circles where we know the people we are talking to aren't just a collection of perfectly rendered pixels. The party tricks are over; the era of the synthetic siege has begun.

📧 Grow Your Audience: Build a newsletter that pays with Ghost — Free for creators