Why This Is Different Now
Deepfakes are built by training a generator model against a discriminator that keeps rejecting its fakes until it can’t tell them from the real thing.
The practical lesson is that any defense that depends on a person spotting visual flaws has a shelf life. And with the explosion of AI and LLMs, that shelf life keeps getting shorter and shorter.
“Something You Are” Is Now Something an Attacker Can Render
Multi-factor authentication (“MFA”) combines at least two of the following elements:
- something you know (such as a password)
- something you have (such as a token)
- something you are (such as a fingerprint)
The “are” factor, whether a face, a voice, or another physical trait, has always been attractive because it is convenient and, well, you always have it with you! Deepfakes attack exactly that convenience as they can defeat video and voice verification. This begs the question that every identity team should be asking: do we need to reconsider using an individual’s likeness as a primary identifier?
Two weaknesses stand out:
- Human-performed checks are the softest target. “Turn your camera on so I can see it’s you” was once a reasonable control. The support desk person squinting at a screen may not easily be able to detect a real-time face swap or a cloned voice.
- Likeness can’t be rotated. A leaked password can be changed. A face and a voice cannot, and executives and staff leave plenty of sampling fodder in conference talks, earnings calls, and social media. These remnants can be pieced together to construct what sounds like actual verbiage or commands from the compromised individual.
Detection tools are not a clean fix either. These include facial artifacts, audio-visual mismatch, pixel-level and motion analysis, metadata and hash analysis, gaze tracking, lighting discrepancies, and temporal artifacts. It also flags two persistent issues: availability and false positives.
The recently published PCI information supplement on authentication guidance points toward layering. It recommends reassessing authentication controls for impersonation risk and adding supplementary signals for assurance: device or computer ID, location, operating system, access time or type, and typing behavior. One caveat matters: those signals do not independently satisfy many MFA requirements or PCI DSS MFA requirements. They add confidence to a real factor. They do not replace one.
It is important to treat a face or voice match as one input, never the deciding one, and keep at least one factor an attacker can’t generate from public media, such as a hardware-backed authenticator.
The Helpdesk: Where a Familiar Voice Becomes a Master Key
A convincing face or familiar voice could impersonate an employee or executive, bypass identity-verification processes, and request account changes. Nowhere is that more dangerous than the helpdesk. A password reset, and especially an MFA re-enrollment, hands an attacker every factor without ever having to defeat one.
Picture a common scenario. The helpdesk gets a call from a director who is traveling, has lost her phone, and is about to present to the board. The voice is right, because it was cloned from a recorded webinar. The caller knows her employee ID and employment start date, which are easy to find. The urgency and emotional pressure are exactly what the deck lists among social engineering warning signs. The analyst may even be invited to a quick videoconference call where the director’s face appears on camera. “I know this person” is the instinct being exploited.
This is not hypothetical for video. There is the Arup case, in which a Hong Kong employee was duped by an AI-generated video call into sending cash to criminals, in a scam reported at $20 million. Mainstream Teams and Zoom video filters have effectively normalized a slightly “processed” look on a call reducing suspicion.
What can your helpdesk do to help reduce this risk?
- Stop counting recognition as verification. A familiar voice or a face on video should carry no weight, and that includes executives. Especially executives!
- Verify out of band. Call back on the number on file or require approval through an already-enrolled authenticator or registered device.
- Tier by risk. MFA resets and privileged-account resets should need manager or security approval through a separate channel.
- Add deliberate friction. Use a cool-down period, notify all existing devices and contacts, and limit access immediately after a reset.
- Re-evaluate Helpdesk challenge questions. AI tools can guess answers effectively with social media searches or other searchable background information.
- Use the supplementary signals. Flag resets that come from an unfamiliar device, location, or time, keeping in mind these support the decision and don’t make it.
- Give staff permission to say no. Scripted refusals and a no-VIP-bypass policy protect analysts from being socially engineered into a fast “yes.”
- Include the latest Deepfake threat landscape in security awareness training. A good start would be to include deep fake simulations similar to phishing simulations as part of the general security awareness training.