Guides
Deepfakes and AI-driven fraud: what cyber and crime policies cover
How AI voice and video impersonation is used to move money, which policy answers a deepfake loss (crime, cyber or neither), and the controls insurers expect.
A deepfake fraud is usually a social engineering loss with a better disguise: someone who looks and sounds like your CFO or a vendor persuades an employee to send money or credentials. Insurance for that loss sits mostly in a crime policy's social engineering endorsement or a cyber policy's funds transfer fraud coverage, both of which tend to carry small sublimits and a verification condition, and neither was written with AI video in mind. Whether a particular policy pays depends on its wording, so confirm it before a loss rather than after.
What the numbers show
The FBI's Internet Crime Complaint Center (IC3) reported in its 2025 annual report that it received 22,364 complaints with an AI reference, with adjusted losses of about $893 million. Most of that was investment fraud, where losses with an AI link topped $632 million. Business email compromise (BEC) complaints with an AI link totaled about $30 million, against $3.05 billion for BEC overall. The FBI cautions that the AI figure understates the problem, because many victims never learn how much AI was involved. The report describes AI-generated emails mimicking a CEO and voice cloning used to request wire payments.
The case that made deepfake video fraud real for finance teams was the Hong Kong office of the engineering firm Arup. Hong Kong police said an employee was convinced by a video conference in which the other participants were fabricated, and sent HK$200 million (about US$25 million) in 15 transfers. Arup confirmed it was the victim in May 2024, as reported by CNN. The lesson for a finance department is that seeing and hearing a familiar face on a call is no longer verification.
How the fraud works
- Executive impersonation. A cloned voice or live video of an owner, CFO or controller instructs an employee to wire funds or change a vendor's bank details.
- Vendor and customer impersonation. A spoofed or hijacked email thread, now backed by a convincing phone call, redirects a payment.
- Help desk and credential attacks. A caller with a cloned voice persuades an IT help desk to reset a password or enroll a new multi-factor device, which can start a ransomware event instead of a wire.
- Fake candidates and fake endorsements. The IC3 report describes voice and video spoofing in job interviews, aimed at gaining network access, and AI-generated celebrity or CEO videos in investment scams.
- Reputation attacks. A fabricated clip of an executive or an employee makes false statements about your products.
Which policy answers
| Loss | Policy to look at | What decides it |
|---|---|---|
| Employee wires money after a fake executive call | commercial crime insurance: what it covers and what it excludes with a social engineering endorsement, or cyber funds transfer fraud | Sublimit, whether the policy names impersonation by phone and video, and the verification condition |
| Spoofed vendor email changes payment details | Same | Whether the loss resulted "directly" from the fraud; see below |
| Attacker uses a cloned voice to gain network access, then deploys ransomware | Cyber first-party coverage (first-party vs third-party cyber coverage) | The ransomware sublimit and whether the intrusion method matters |
| Fake video of your executive harms the brand | Cyber with a reputation or deepfake endorsement; sometimes nothing | Whether a "deepfake event" is a defined trigger |
| Lawsuit by someone harmed by AI-generated content you published | Liability policies | New ISO generative AI exclusions; see AI liability insurance: how insurers are treating artificial intelligence risk |
The crime policy and the word "directly"
Crime forms pay for loss "resulting directly from" a covered act, such as a fraudulent instruction. Traditional computer fraud and funds transfer fraud insuring agreements were drafted for hackers and forged bank instructions, not for an employee who authorizes a payment after being deceived. Courts reached different results in spoofed-email cases: the Second Circuit in Medidata v. Federal Insurance (2018) and the Eleventh Circuit in Principle Solutions v. Ironshore (2019) found coverage, while the Sixth Circuit in American Tooling Center v. Travelers (2018) did not, because an employee's own approval intervened. Those cases involved email, not AI video, but a deepfake raises the same causation question, and your state's law and the exact wording will decide it. That uncertainty is why carriers sell a separate social engineering endorsement that covers voluntary payments by deceived employees. See social engineering fraud: cyber policy or crime policy? and business email compromise and funds transfer fraud coverage.
Those endorsements usually carry a sublimit well below the crime policy's main limit, and many require that the employee verified the instruction through a known contact method before paying. A callback to a number supplied by the person asking for the payment is not verification. Neither is a video call.
The cyber policy
Cyber forms often include funds transfer fraud or social engineering as a sublimited first-party coverage, and a few carriers have started to write deepfakes in by name. In December 2025, Coalition announced a deepfake response endorsement for its cyber policies in the United States (on a non-admitted basis) and several other countries, with a forensic analysis of the suspect content, legal help to have it removed from platforms and crisis communications support; its release does not say the endorsement covers fraudulent funds transfers. BOXX Insurance announced in July 2026 an endorsement that adds affirmative coverage for AI and deepfake events tied to social engineering and security failures. Insurance Business reported that carriers have gone in opposite directions since January 2026, with some excluding AI-generated deepfake fraud from social engineering coverage and others adding it. Do not assume the market is converging. Ask what your own renewal says.
A hypothetical: the fake vendor call
Say a 40-person building-products distributor receives an email, apparently from a long-time supplier, saying its bank account has changed. A person who sounds like the supplier's owner confirms by phone, and an accounts-payable clerk wires $180,000. The numbers here are invented to show the mechanics.
- The clerk learns of the problem two days later. The first call is to the bank, to ask for a recall, then to the FBI's IC3 at www.ic3.gov. The IC3 Recovery Asset Team's Financial Fraud Kill Chain, in 2025, assisted with 3,900 incidents and froze about $679 million of $1.16 billion in attempted theft, a 58% success rate, per the IC3 report. Speed is the main variable.
- The distributor notifies the crime carrier and the cyber carrier. Only one may have a social engineering grant, but notice to both preserves rights.
- The crime policy carries a $100,000 social engineering sublimit and a verification condition. The clerk emailed back to the same address and called the number in the email, so the carrier questions whether the condition was met.
- The result turns on that wording, the sublimit and the retention, not on the size of the loss.
Controls that matter to underwriters and to claims
Underwriters ask about these on the application (the cyber insurance application: what underwriters ask and how to answer), and a missed one can become a reason to deny.
- Out-of-band verification. Confirm any change of bank details, and any unusual payment request, by calling a number already in your records.
- Dual authorization for wires above a threshold you set, with different people initiating and approving.
- A vendor bank-change procedure that no executive can waive by phone or video.
- Help desk rules that bar password and MFA resets on the strength of a voice, with a call-back or in-person check.
- A shared code phrase for executive requests, changed periodically.
- Training that uses recorded cloned-voice examples.
The cyber insurance readiness checklist covers the technical controls, and MFA, EDR and backups: the controls cyber insurers now require shows what carriers check.
Questions for your broker
- Does our social engineering coverage name impersonation by voice, video or AI, and does any AI or deepfake exclusion apply?
- What are the sublimit, retention and verification condition, and do they sit in the crime or the cyber policy?
- Are payments by an employee to a fraudulent vendor covered, or only transfers initiated by the criminal?
- Does the cyber policy respond to a reputation attack from a deepfake, and what are the covered services?
- Do crime and cyber coordinate, or do they each point at the other?
Key takeaways
- A deepfake fraud is mostly a social engineering loss; insurance answers through a crime endorsement or cyber funds transfer coverage, both usually sublimited.
- Courts have split on whether a deceived employee's payment is a "direct" loss under older computer fraud wording, so ask for explicit social engineering coverage.
- Carriers are inconsistent on AI: some exclude deepfake fraud, others add endorsements. Read the renewal.
- Verification by call-back to a known number, dual approval and fast bank recall decide both the loss and the claim.