Verbal passphrases and hardware security keys are emerging as the most reliable AI deepfake defence protocols available to organisations, after researchers and government agencies concluded that human detection of synthetic voices and faces has fallen to near-chance accuracy. The shift follows a string of costly incidents that exposed the limits of visual identity verification.
How a $25.6 Million Video Call Exposed the Flaw
In January 2024, a finance employee at professional services firm Arup joined what appeared to be a routine video conference. Every person on the other side was an AI-generated clone, built from publicly available conference footage and virtual company meetings featuring Arup executives. The call ended with 15 wire transfers totalling $25.6 million (HK$200 million) sent to five Hong Kong bank accounts, according to CFO Dive.
Arup did not publicly identify itself as the victim until May 2024, though Hong Kong police had disclosed the incident in February 2024. As of early 2025, none of the funds had been recovered, with Hong Kong authorities still investigating, according to PurpleSec.
‘Seeing and hearing someone is no longer proof they are real,’ said Deepak Gupta, technical chief executive of GrackerAI, when discussing the Arup case. ‘Any protocol that relies on “I recognised their face and voice” is now broken.’
The science supports that concern. A University College London study found listeners could identify deepfakes accurately only about 73% of the time. Even with dedicated training, accuracy improved by just 3.84%. Researchers from the University of Duisburg-Essen and Indiana University later aggregated results from 56 similar studies and found human detection rates were closer to chance.
AI Deepfake Defence Protocols Experts Now Recommend
A joint cybersecurity information sheet released by the NSA, FBI, and the Cybersecurity and Infrastructure Security Agency (CISA) on 12 September 2023 ruled out automated detection tools that scan for visual or audio manipulation artefacts, noting those methods can no longer be relied upon to find statistically significant traces of tampering. CISA and the NSA both recommend that organisations implement real-time verification capabilities alongside passive detection techniques and rehearse responses to exploitation attempts.
Where automated tools fall short, security practitioners are turning to lower-technology controls. James Scobey, chief technology officer at B2B cybersecurity firm S2i2, and GrackerAI’s Gupta both advocate for verbal passphrases and hardware-based security keys. CISA now recommends FIDO2 and PIV hardware credentials as the new standard for multi-factor authentication. The FBI separately advises organisations to use secret verbal passphrases shared among colleagues to verify identity on voice and video calls, per its North Korean IT worker threat guidance.
A verbal passphrase works by requiring the person on a call to confirm a pre-agreed secret word or phrase before any sensitive action proceeds. Scobey warns that attackers routinely manufacture urgency to pressure employees into bypassing checks. His fix is direct: ‘Make the control automatic and no-exception.’
For larger organisations, Scobey recommends treating passphrases with the same discipline applied to standard password management. That means generating them randomly rather than choosing them manually, using unrelated words separated by numbers and symbols (for example, ‘harley9jedi@buddies.sinclair’), assigning separate passphrases to different roles and transaction tiers, and rotating them on evidence of compromise rather than on a fixed calendar. NIST guidance now discourages fixed 90-day resets precisely because they create predictable patterns. High-value transactions should also require out-of-band callbacks and dual authorisation from a second employee before any payment is processed.
North Korean Operatives Raise the Stakes
The threat extends well beyond one-off financial fraud. Security awareness training company KnowBe4 discovered in 2024 that it had hired a North Korean operative for the role of Principal Software Engineer after conducting four video interviews, running a background check, and confirming the applicant’s photo matched identity documents, according to Axios. The candidate had stolen a US-based identity and used AI tools to alter a stock image to pass an ID check.
Once KnowBe4 shipped the operative a Mac workstation, it immediately began loading malware. Anti-malware software flagged suspicious activity on 15 July 2024 at 9:55pm EST; the workstation was contained approximately 25 minutes later after the employee failed to answer a phone call from the IT team, according to SecurityWeek. The operative used a Raspberry Pi to download malware and attempted to manipulate session history files, per CyberScoop. KnowBe4 confirmed on its official blog that no data was accessed, compromised, or exfiltrated.
The incident is not isolated. The US Department of Justice determined that North Korean actors fraudulently obtained remote IT roles at more than 100 US companies, assisted by collaborators in the United States, China, the United Arab Emirates, and Taiwan.
‘The defence against the most advanced AI attack is often deliberately low-tech,’ Gupta said. Deepfake technology can replicate anything observable in public channels, but it cannot replicate a randomly generated secret that exists only between two people in a controlled, private exchange. Organisations that tie passphrase authentication directly into finance platform workflows, rather than relying on employees to remember the protocol under pressure, will face the stiffest test of whether the controls hold when an attacker next calls posing as the chief financial officer.