National Cybersecurity
Association of Ukraine

The most expensive attacks are not technical: what the loss data shows.

Guide G14 · Published

For executives, finance teams and anyone who signs off on money. Rank attacks by the money lost rather than by the number of incidents, and the most expensive ones turn out not to break software at all. They break the procedure by which a person approves a payment.

Who this is for and when to use it

This is not a threat overview or a tool list. It answers one management question: where to put attention when the security budget is finite. The answer comes from recorded losses rather than from forecasts.

The figures below are from the 2025 annual report of the FBI's Internet Crime Complaint Center 1. It is the largest open dataset of reported cybercrime losses, and it has known limits, which have a section of their own at the end.

What the figures show

In 2025 IC3 received 1,008,597 complaints, and reported losses passed twenty billion dollars for the first time, reaching $20.877bn, 26 per cent more than the year before. The average loss per complaint was $20,699 1.

The interesting part is the ranking. Sorted by money lost, the top crime types are 1:

  • investment fraud - $8,648,617,756
  • business email compromise (BEC) - $3,046,598,558
  • tech and customer support fraud - $2,134,675,818
  • personal data breach - $1,314,923,988
  • confidence and romance fraud - $929,287,469
  • government impersonation - $797,943,193

And here is the reason this guide exists. Ransomware, the category most security conversations are built around, accounted for $32,320,105 of reported losses in the same year 1.

Business email compromise cost victims roughly 94 times what ransomware did. BEC needs no software vulnerability at all: it needs one message that somebody believed.

The fastest-growing category is not technical either. Government impersonation complaints rose from 17,367 in 2024 to 32,424 in 2025, close to doubling in a year 1. Complaints flagged as AI-related numbered 22,364 1.

Why this is so

Breaking a modern operating system is expensive. Finding a vulnerability, keeping it quiet and evading detection is skilled work that takes time. Persuading an accountant to change the bank details on an invoice costs one email.

So the attack moved to where there are no patches. It targets the procedure rather than the code: the habit of trusting a familiar signature, the reluctance to double-check with a manager, the rush at quarter end, and the fact that the request for a payment and the confirmation of it arrive through the same channel.

Generative tools widened that gap. Synthetic voice and video made convincing what used to be checked by ear. National cyber agencies now say plainly that checking where a piece of media came from belongs to basic hygiene rather than to specialist forensics 3.

No technical measure settles the question on its own. NIST's overview of approaches to digital content transparency sets out both what such measures give you and where they stop 4. A provenance signature does not make a file true, and its absence does not make a file fake.

What to do about it

The practical conclusion is that the largest gain comes from a written confirmation procedure rather than from new software. The minimum below costs nothing.

  1. 01Adopt a second-channel rule: any change of bank details, and any payment above a set threshold, is confirmed through a different channel than the one the request arrived on.
  2. 02Write down that the number used for confirmation comes from your own internal directory, never from the message or its signature block.
  3. 03Give people the right to pause. State that an employee is not answerable for a delayed payment if they stopped to verify. Without this the procedure fails, because urgency is the main lever attackers use.
  4. 04Agree a family or team code word for phone requests about money. Against a cloned voice this works better than recognising someone by ear.
  5. 05Rehearse the scenarios once a year: a fake request from a manager, a fake change of supplier details, a fake call from the bank.

Detailed protocols for individual scenarios are in our guides on voice cloning, on fake executive video and on lookalike domain names.

Limits of this data

These figures should not be read as a picture of the world, for four reasons.

  • They are complaints made to a United States agency. Most complainants are American, so the proportions may differ elsewhere.
  • They are reported losses, not all losses. A large share of incidents is never reported, particularly when the victim is an individual or a small company.
  • Money lost and incident count are different measures. By count the largest category is phishing and spoofing, 191,561 complaints 1, although its direct losses are far below those of BEC.
  • A low ransomware loss figure does not mean ransomware is harmless. These totals exclude downtime, recovery and reputational cost, as the report itself notes.

This guide is about where attention goes, not about dropping technical defences. Backups, updates and multi-factor authentication remain mandatory.

Sources

  1. 1FBI Internet Crime Complaint Center (IC3). 2025 Internet Crime Report · accessed 2026-11-18
  2. 2FBI Internet Crime Complaint Center (IC3). Annual Reports index · accessed 2026-11-18
  3. 3Canadian Centre for Cyber Security, with NSA, Australian Signals Directorate's ACSC and UK NCSC. Joint guidance: Content Credentials and strengthening multimedia trust in the generative AI era · accessed 2026-11-18
  4. 4NIST. Reducing Risks Posed by Synthetic Content: an overview of technical approaches to digital content transparency (NIST AI 100-4) · accessed 2026-11-18