Here’s what AI attacks actually look like

Discussions about AI-powered attacks happen behind closed doors, tucked away in internal incident reports and confined to candid exchanges between peers at industry conferences. This secrecy leaves security leaders with an near-impossible task: responding to threats they can’t see.

That’s why we’re breaking down the anatomy of AI attacks: what they look like, how they work, and why they’re able to slip past traditional defenses.

Here are the stakes: A multi-million dollar wire transfer after a deepfake video call.¹ A North Korean operative moving through your hiring pipeline.² An IT helpdesk handing over SSO credentials to someone posing as an employee.³

Attacks like this are playing out at companies everywhere, everyday. And with AI-driven attacks growing by nearly 14x in six quarters,⁴ organizations can’t afford to treat this as low-risk threat.

The scams are here. Is your security strategy ready for them?

AI has forced security leaders to doubt the channels they trust

Talk to a CISO right now and one word keeps coming up: uncertainty. Uncertainty about how AI attacks will reshape security strategies. Uncertainty about where to invest. Uncertainty about where the real risk even lives.

That’s exactly why we created the CISO Deepfake Defense Council.The Council brings together a select group of highly regarded security executives from seven Fortune 500 enterprises and other category-leading organizations across a broad cross-section of industries, including technology, financial services, healthcare, life sciences, defense, enterprise software, cryptocurrency, and cybersecurity. The Council is focused on delivering strategic guidance to help enterprises understand, prepare for, and defend against deepfakes—one of the fastest-emerging threats to trust and identity.

We interviewed these leaders individually and as a group, probing to understand how they’re thinking about the new enterprise threat landscape.

Takeaways from the CISO Deepfake Defense Council

How CISOs think about the new threat landscape

Council members mapped AI attacks by reputational and financial risk. Answers varied, but six threats rose to the top.⁵ This briefing covers those six.

Where do these attacks land on the risk exposure grid for your organization? Download a blank version and bring this exercise back to your team.

Your applicant pool is full of potential insider threats

In 2025, the FBI released an official public service announcement warning about North Korean operatives trying to get hired at U.S.-based companies.¹ These operatives use fabricated or stolen identities to apply for jobs, funneling salaries back to the regime while conducting espionage from inside company walls.

Deepfake applicants show up with polished resumes, credible LinkedIn^® profiles,² and impressive work history, and when they log on for an interview, they seem well-prepared and professional. But these fake candidates often use real-time face-swap tools trained on photos, sometimes scraped from real people’s social media, filtered or manipulated voices, and work references that lead to companies that don’t exist or can’t be verified.

The consequences of hiring a fake candidate can be severe. Once inside a company’s systems, a fraudulent employee can steal data and intellectual property, or plant backdoors for future exploitation. Companies that unknowingly pay these workers may also find themselves in violation of federal sanctions laws. And being infiltrated by a state-sponsored actor can erode customer trust and bring unwanted scrutiny.

And when it works, the damage hits fast. Funds are drained before the customer even knows their account has been touched. For the organization, that means fraud losses, costly incident reviews, and reputational damage. In an industry where trust is everything, a wave of successful account takeovers can do lasting damage.

Weaponizing authority for deception

When an employee gets a suspicious email, security training tells them to verify through a different channel. But when they join a video call and see a face they recognize, that feels like verification. Even though it isn’t. In the AI era, seeing and hearing someone is no longer proof that they’re real, and most employees don’t know that yet.

Executive impersonation is happening now, and virtual meetings are one of the biggest blind spots enterprises have. A cloned voice or face in a meeting carries the same authority as the real thing, and in many cases, nobody in that room can tell the difference. A random text asking for gift cards is easy to dismiss, but a call that sounds and looks identical to your CEO? That’s much harder to ignore.

Deepfake leaders can end in multi-million dollar mistakes

Think about what happens when an employee joins a meeting with their CEO who asks them to move fast on a wire transfer. The instinct is to help, to not slow things down, and to trust the person at the top. Executive impersonation works because it exploits something security training struggles to override: the instinct to comply with authority.

And the damage goes beyond manipulated employees. Fake videos of leaders making statements about financials or product updates can do serious damage to brand reputation too. Leaders have access, influence, and trust. AI lets scammers mimic all three.

Real stories and real damage of fake leaders

How helpdesk attacks play out

Armed with basic personal data from social media and leaked credentials, most of it available for mere dollars on the dark web, attackers call in, impersonate an employee, and request an MFA reset. If the helpdesk relies on knowledge-based authentication alone, they open the door for the scammer. From there, it’s a straight path to unauthorized access, ransom, IP theft, and the ability to move anywhere in your systems.

That was all possible through impersonation at the helpdesk.

With AI, attackers can convincingly impersonate an individual’s voice and even face to aid in their scheme. They can also use AI to automate attacks, probing at channels like the helpdesk to find vulnerabilities to exploit, turning this once-trusted channel into a wide open door for attackers.

Victims are tricked by a relationship they believe is real

Bad actors build fictitious advisor or client personas, complete with robust LinkedIn^® profiles,¹ fabricated credentials, and believable networks, and use them to infiltrate investor transactions. Once credibility is established, the ask follows: invest here, move funds, change account details, or redirect a wire.

This works because digital trust in high-value financial relationships is still built on perception rather than verification. A credible profile, the right language, and a warm introduction are enough.

People with authority and expertise like Rosenberg have influence—which makes them a prime target for impersonation. AI lets these attackers impersonate either advisors or high-net worth individuals to steal funds with ease.

Third-party relationships are invaluable—and easily manipulated

Vendor impersonation and partner impersonation are some of the most underestimated risks enterprises face right now. These are relationships built on trust, and they operate across external systems that are inherently harder to monitor and verify. Unlike internal communications, the channels aren’t as secure. An attacker posing
as a trusted supplier just needs to send a realistic email, a cloned voice message, or a fake invoice to break through defenses.

A financial institution (and Pindrop customer) believed initially they only had a problem with deepfakes in hiring. After only deploying Pindrop Pulse^® for meetings to their hiring team, an employee outside their recruiting team flagged a supposed “vendor” as suspicious. They asked for access to Pulse and, within seconds, caught a deepfake.¹

For attackers, AI is becoming the force behind their operations

A single check at login means nothing when attackers can generate a convincing identity, voice, and face in real time, and enter the channels you trust. The answer is continuous identity verification: ongoing, multi-signal validation throughout every real-time interaction, every call, every remote interview, every vendor communication.