The best deepfake defense may be a secret phrase

A face and voice on a live video call are no longer reliable proof of identity. ZDNET reports that security specialists are advising companies to bring back deliberately low-tech controls—including secret verbal passphrases—for high-value requests that increasingly face AI-powered impersonation.

As first reported by Zdnet.

The risk became clear in January 2024, when an employee at professional services firm Arup joined a video call with people who appeared to include the company’s chief financial officer. The participants were AI-generated clones assembled from public appearances and earnings calls involving Arup executives. The call led to 15 wire transfers totaling about $25 million to third-party accounts.

Recommended reading

Apple fixes 20+ iPhone security flaws in iOS 26.6.1

Sergey Kuznetsov • Aug 17, 2026 • 2 min read

“Seeing and hearing someone is no longer proof they are real. Any protocol that relies on 'I recognized their face and voice' is now broken.”

Why visual and voice checks are failing

Live calls have traditionally served as an informal identity check for sensitive financial transactions, medical information, and legal discussions. Employees often relied on visual cues, familiar voices, background sounds, or signs such as unnatural breathing and synthetic modulation to spot an impostor.

Those signals are becoming less dependable. A University College London study found that listeners correctly identified deepfake audio and video only about 73% of the time. Training and familiarity with deepfake examples improved that result by just 3.84%. A separate aggregation by researchers from the University of Duisburg-Essen and Indiana University examined 56 similar studies and found that human detection rates were closer to chance.

James Scobey, CTO at B2B cybersecurity firm S2i2, said conventional tells can still provide a supporting signal, but should not be treated as an effective control. The danger is that training workers to trust their perception puts them at the center of an attack designed specifically to manipulate that perception.

A joint information sheet from the NSA, FBI, and Cybersecurity and Infrastructure Security Agency (CISA) also rejected the assumption that automated detection tools can always identify manipulation in voice or video. Such systems depend on finding statistically meaningful traces of editing or synthesis, but increasingly capable fakes may not leave a consistent signature.

The threat is also moving beyond one-off payment scams. In a 2024 incident, security training company KnowBe4 hired an attacker who passed interviews and screening, then provided the person with a company workstation. The company later discovered that the worker was a North Korean operative using the device to upload malware into its systems.

Gupta said North Korean attackers have been operating similar schemes at scale, sending thousands of fake workers each year to pose as employees at companies in the US and Europe. A 2025 notice from the US Department of Justice said some operations had assistance from collaborators in the US, China, the United Arab Emirates, and Taiwan. Google’s Threat Intelligence Group has investigated related incidents since 2022, with a significant portion now targeting European countries.

How verbal passphrases and hardware keys work

A verbal passphrase is a secret word or phrase shared privately among authorized members of a company or household. During a voice or video call, the recipient asks the caller to provide the phrase. Because the information is not normally available through public appearances, earnings calls, or other observable channels, a deepfake built from public material cannot simply reproduce it.

The control only works if employees use it every time. Attackers commonly create urgency or exploit workplace hierarchy, persuading employees to skip checks when the caller claims to be a senior executive. Scobey said the requirement should therefore be automatic and allow no exceptions.

“Make the control automatic and no-exception.”

Scobey also recommends simulated deepfake and voice-phishing calls during employee training. The goal is to give workers practice refusing an urgent request from someone who appears to hold authority, rather than expecting them to identify every manipulation by sight or sound.

For larger organizations, the proposed approach includes:

• Generate passphrases with a random password generator instead of choosing them manually.
• Use unrelated words separated by random numbers and symbols. One example from Scobey is harley9jedi@buddies.sinclair.
• Maintain separate phrases for different roles and transaction ranges, rather than creating one master secret.
• Replace a phrase after evidence of compromise, a role change, or employee offboarding—not according to a predictable fixed schedule.
• Combine the phrase with hardware keys, out-of-band callbacks, and dual authorization.

Scobey pointed to the Electronic Frontier Foundation’s word list as a source for constructing memorable random strings, but said randomness matters more than meaning. Human-selected phrases tend to reflect subconscious preferences and are therefore easier to guess.

Role- and transaction-specific phrases can also limit the damage from a compromise. For example, an accounts-payable team might use one phrase for transactions below $1,000 and another for transactions from $1,000 to $20,000. If one phrase is exposed, it can be replaced without invalidating every other authorization path.

Passphrases are not enough for high-value payments

A verbal phrase should not stand alone for major enterprise transactions. Security teams can require an out-of-band callback through a separate official channel and a second authorized employee’s approval before money moves. That creates independent checks instead of relying on the same compromised call.

CISA recommends FIDO2 and PIV hardware credentials as the new standard for multifactor authentication, while FBI guidance has repeatedly included secret verbal passphrases. A password manager authorized for federal use, such as Keeper Security, can help organizations manage the resulting collection of credentials and secrets.

The tradeoff is additional friction. Companies often resist adding steps to workflows, but Gupta argues that controls should be concentrated around high-risk actions rather than imposed on every routine task:

“When people experience security as targeted rather than blanket, they stop trying to bypass it.”

The operational shift is straightforward: stop asking employees to prove that a voice or face looks authentic, and require the requester to pass a secret, device-based, or independently verified control instead. Deepfake detection can remain a supporting signal, but the payment should depend on credentials the attacker cannot obtain from a public video.