AI Deepfake Social Engineering: How Enterprises Can Defend Against It

Watch the full session · 27 min

Deepfake fraud attempts have surged 21,000% in three years, according to TrendAI Global Deputy CISO Sundari Parekh. At COLLIDE 2026, she walked through the Arup and Ferrari cases, explained why defenses must run at machine speed, and listed the actions security teams should take in the next 90 days.

Deepfake fraud is no longer a crude text asking for gift cards. Attackers now copy a real executive's voice, face and mannerisms, Sundari Parekh, Global Deputy CISO at TrendAI, told attendees at COLLIDE Data + AI Conference 2026 in Atlanta.

In her session, "AI-Driven Social Engineering: Defending Against the Next Phase of Cybercrime," Parekh described how AI agents are speeding up social engineering attacks and laid out the defenses she recommends security teams put in place now.

How are attackers using AI for social engineering?

Parekh said deepfake fraud attempts have surged 21,000% in the last three years, which she said works out to one deepfake every five minutes in 2024. She added that almost half of companies are being targeted by deepfake fraud.

The old model was a text claiming the CEO urgently needed $500 in gift cards. Today, she said, attackers build a persona from public material. As a public-facing security leader, her voice and image are easy to find online, so someone could impersonate her and order a colleague to act during a fake security incident. "The deepfakes are getting a lot more targeted," Parekh said.

AI also speeds up each stage of an attack, she said. Agents research targets on LinkedIn, websites and forums, then execute and adapt in real time. Finding a company's data and financial systems once took time. "Now it takes a few minutes to do that," she said.

What can enterprises learn from the Arup and Ferrari deepfake cases?

Parekh pointed to two cases. At Arup, she said, an AI-generated CFO on a video call authorized 15 transactions, and the company lost $25.6 million. "The video call creates a false sense of legitimacy," she said. When a familiar face appears on a sanctioned collaboration tool, people trust it.

In a near miss at Ferrari, an attacker used a cloned CEO voice on a WhatsApp call to push an urgent deal. The attack was halted when verification questions couldn't be answered. The lesson, Parekh said, is that these attacks are sophisticated but not perfect, and human verification still works.

Her conclusion was that verification protocols, including two-step verification and authorization for financial transactions, are essential. She encourages her own team to call or email her back, or check with the CISO, whenever a request sounds strange.

How do you defend against AI attacks at machine speed?

Traditional, reactive security cannot keep up with the volume and speed of agentic attacks, Parekh said. Platforms have to process millions of events per second, and automated threat detection and response can no longer be optional or outsourced as an afterthought.

"AI-powered defense mechanisms are essential to AI-powered attacks," she said. Human analysts move up to strategic decisions. "Gone are the days of the SOC 1 analyst," Parekh said, adding that an agent can now do that work while humans stay in the loop.

She also described TrendAI's newly announced autonomous vulnerability discovery, which uses agentic AI with the company's Zero Day Initiative bug bounty program to find exploitable vulnerabilities. She compared vulnerabilities to unlocked windows in a house: attackers check each one until they find an opening, and defenders need to know which windows are open and protect them until they can be fixed.

What should security teams do in the next 90 days?

Parekh listed several immediate actions:

• Use AI and machine learning for threat detection across infrastructure, with real-time monitoring and automated response for high-confidence threats.
• Run security awareness training on AI threats and deepfakes. Employees should feel free to pause on urgent requests, even from a manager or VP. "We have a gut for a reason, and just validate it," she said.
• Write an incident response playbook for AI-powered, multi-vector attacks, with clear roles. During an incident, she said, "everybody wants to play the hero cowboy," and that is not what's needed.
• Build behavioral analysis and anomaly detection for agents, similar to "impossible travel" alerts for user logins, so misbehaving agents can be shut down.
• Strengthen authentication beyond passwords and standard MFA. She urged teams to adopt passwordless login.
• Architect redundancy and failover, assuming primary defenses will be compromised, including air-gapping critical data.
• Work with threat intelligence communities and law enforcement, such as forums run by the FBI, CISA and NSA.

On resilience, she described a fellow CISO whose company lost its order management system to ransomware and rebuilt in the cloud but lost its order history. "In the age of AI, we have to think more about resilience than we ever did before," Parekh said.

Her closing takeaways: AI agents are accelerating criminal capabilities today, traditional defenses can't match machine-speed attacks, automated security is non-negotiable, and continuous preparation is a competitive advantage.

AI security sessions at COLLIDE 2027

COLLIDE, the largest AI conference in the Southern U.S., returns to Atlanta on Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center, bringing together more than 1,000 enterprise data and AI leaders. Learn more about COLLIDE, or find upcoming sessions in Data Science Connect's free webinar series.

Watch: AI-Driven Social Engineering: Defending Against the Next Phase of Cybercrime on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights