Marvel Advisors
Contact us
AI-Powered Attacks

AI-Powered Attacks: Phishing, Deepfakes, and What Changes for Incident Response

Oct 29, 2026 ยท By Marvel Advisors

For years, security awareness training taught people to spot phishing by its mistakes: bad grammar, odd phrasing, a generic greeting. Generative AI removes most of those tells. It also makes it practical to imitate a specific person's voice, and sometimes their face, on a call. The underlying scams are not new. What has changed is how convincing and how cheap they have become.

What the data shows

The FBI's 2025 Internet Crime Report is the most useful public benchmark:

  • Business email compromise (BEC) generated 24,768 complaints and about $3.05 billion in reported losses in 2025.
  • IC3 received 22,364 complaints that included AI-related information, with adjusted losses of more than $893 million.
  • Businesses reported more than $30 million in losses to BEC scams involving AI.

Read those AI figures as a floor. They count only complaints that included AI-related information, and the report itself observes that many victims do not realize the extent AI may be involved.

How attackers use AI

Phishing and BEC. The FBI describes chat generators being used to produce official-sounding emails that mimic a CEO or other officials, containing phishing links or wire instructions. AI also makes it easy to tailor messages to each target using public information and to write fluently in any language.

Voice cloning for payment fraud. The same report notes that voice cloning can be used to request wire payments. Anyone whose voice is public, in webinars, podcasts, or videos, is a potential source for a cloned request "from the CFO."

Impersonation campaigns. In May 2025 the FBI warned of an ongoing campaign using text messages and AI-generated voice messages that claimed to come from senior U.S. officials, and noted that AI-generated voices can sound nearly identical to the real person.

Help-desk social engineering. CISA's advisory on the Scattered Spider group (updated July 2025) describes attackers posing as employees and calling IT help desks to reset passwords and transfer MFA to devices they control. Voice and video deepfakes make this kind of call harder to challenge.

Fake candidates. The IC3 report also describes voice spoofing, or potentially voice deepfakes, in online job interviews, where the goal often appears to be gaining access to company networks rather than a direct payment.

Verification has to move out of the channel

The common thread: you can no longer trust a voice, a face, or a well-written email as proof of identity. Verification has to happen through a separate channel the attacker does not control. The FBI's guidance in its May 2025 alert and December 2024 alert on generative AI fraud points the same way: call back using a number you look up yourself, and independently verify any request for money.

In practice:

  • Payments and banking changes. Any new payee, change to bank details, or urgent wire requires a callback to a number already on file, never one supplied in the request. Require a second approver above a set threshold.
  • Help desk. Do not reset passwords or MFA based on a phone call alone. Verify through the employee's manager, a ticket raised from an authenticated session, or an in-person or video check against an ID on file. Have the user approve the change on a device that is already enrolled.
  • Executives. Agree in advance that leadership will never ask for payments, gift cards, or credentials through chat, text, or voice alone, and tell staff it is always acceptable to verify.

Stronger authentication reduces the damage

If a deepfake does convince someone, the next control is authentication that cannot be talked out of a person. CISA recommends phishing-resistant MFA, such as FIDO/WebAuthn or PKI-based methods, noting in the same advisory that these are not susceptible to push bombing or SIM swap attacks. Prioritize email, VPN and remote access, and administrator accounts.

What changes for detection and incident response

AI-enabled attacks often look like normal business activity: a legitimate login, an approved payment, a routine password reset. That shifts detection toward behavior:

  • Watch identity events closely. MFA device changes, password resets followed by new-location logins, new inbox forwarding rules, and new OAuth app consents are early signs of account takeover.
  • Correlate across email, identity, endpoint, and network. A managed extended detection and response (XDR) or network detection and response service is built for this correlation, so that a help-desk reset and an unusual data transfer an hour later are seen as one incident rather than two unrelated alerts.
  • Update your playbooks. Add specific playbooks for suspected deepfake payment requests and fraudulent help-desk resets. For payment fraud, speed matters: contact your bank immediately to request a recall, and report to IC3.
  • Preserve evidence. Keep the voicemail, recording, email headers, and call logs. They help investigators and insurers.
  • Rehearse. Run a tabletop exercise built around a cloned executive voice. It is the fastest way to find out whether your callback rules hold up under pressure.

Quick checklist

  • Require out-of-band callbacks to known numbers for all payment and bank-detail changes
  • Require dual approval for wires above a set threshold
  • Set a help-desk identity verification procedure that does not rely on voice alone
  • Deploy phishing-resistant MFA for email, remote access, and admins
  • Alert on MFA changes, password resets, inbox rules, and new app consents
  • Add deepfake and help-desk fraud playbooks to your incident response plan
  • Run a deepfake tabletop exercise with finance, IT, and leadership
  • Update awareness training: polished messages and familiar voices are not proof

Being ready when it happens

No control stops every convincing request, so the goal is to catch it quickly and respond well. An incident response plan with AI-specific playbooks, paired with continuous monitoring through managed XDR and network detection and response, shortens the time between "that call seemed off" and containment.

Sources