Yes, but not in the way most coverage suggests. The cons are the same. The two clues people relied on, bad writing and a familiar voice, no longer hold.
Short answer: Yes, but the change is narrower than the coverage suggests. The scams themselves are the same ones from a decade ago: fake invoices, impersonated executives, romance and investment cons. What has changed is that two of the informal tells people relied on, clumsy English and a recognisable voice, no longer work. The con is old. The polish is new.
Key takeaways
Three things, concretely.
Writing quality. Awkward English was never a security control; it was a lucky by-product of scams being written cheaply by people working in a second language. Generative writing removed it. A fraudulent message can now be as fluent as a real one, in your industry's vocabulary, matching your correspondence style.
Volume of tailoring. Personalising a message used to cost human time, so it was reserved for high value targets. Automating the research and drafting means the tailored version is now affordable against a small business, not just against a corporate finance department.
Voice. Cloned audio from a short sample means hearing a familiar voice is weaker evidence than it was. Reporting indicates scammers have cloned the voices of family members, executives and public figures from small amounts of audio taken from social media.
More than people expect, and this is the reassuring half.
The underlying cons are the same short list. Someone wants you to move money, hand over credentials, or send identity documents, usually with urgency and usually outside your normal process.
The defences are the same too, because none of the good ones ever relied on spotting a mistake. Verifying a bank detail change by phone on a number you already had works identically whether the email was written by a person or a model. Multi-factor authentication does not care how convincing the login page was. A payment approval threshold is indifferent to prose quality.
If your security posture depended on scams being obvious, it has weakened. If it depended on process, it has not.
Australians reported more than 200,600 scams to Scamwatch during 2025, with losses reported to that service exceeding $330 million. Combined reported losses across Scamwatch, ReportCyber, IDCARE, the Australian Financial Crimes Exchange and ASIC were around $2.18 billion for the year.
Reporting through 2026 has described a pattern worth understanding: the number of reports has come down while average losses per victim have risen. Fewer people are falling for scams, and those who do lose more. That is consistent with attacks being better targeted rather than more numerous.
ASIC issued a public warning in 2026 that scammers are using artificial intelligence to build extensive networks of deception, including cloned websites and deepfake identities designed to resemble legitimate organisations.
On voice cloning specifically, one widely reported estimate from an academic at Swinburne University of Technology put Australian losses at about $25.8 million in the first half of 2025. Treat that as an informed estimate rather than an official statistic; the agencies do not publish a separate voice cloning category.
We have deliberately not repeated the very large percentage increases circulating in vendor marketing. Where we could not trace a figure to an agency or a named researcher, it is not on this page.
Four practical adjustments, none of them expensive.
Stop teaching people to spot bad English. If your induction still says look for spelling errors, it is training a reflex that no longer fires. Replace it with: any request involving money or identity documents gets verified, regardless of how it reads.
Treat voice as unverified. A call that comes to you proves nothing. A call you place to a number you already had proves a great deal. That asymmetry is the whole control, and it survives cloning.
Agree a family and staff check. A short agreed phrase, or a question with an answer that is not on social media, defuses the urgent call from a distressed relative or a stranded director.
Reduce the raw material. Tailored attacks are built from true details about your business. The less of your client information, pricing and internal language sitting inside third party services, the less there is to build with.
That last point is where Prinn works, and only that point. It runs on each person's computer and checks a prompt before it is sent, so identity numbers, financial details and credentials are less likely to accumulate inside a tool you do not control. If it finds something sensitive it shows what it found and offers three choices: redact it, stop, or send it as is. The person decides, and it never blocks anyone. Prompt content never leaves the machine, including to us.
Prinn does not detect scams, does not filter email, does not screen calls and cannot tell you whether a voice is real. Anyone selling you AI protection against AI scams is describing a different product than this one.
It is fair to say scams have become harder to spot and unfair to say we are defenceless. The change is real but bounded: better prose and better audio, aimed at the same short list of cons.
The reason to stay calm is that the effective defences were never about detection. They were about verifying the two things that cost you money, payments and disclosures, through a channel the attacker does not control. That was good advice before this technology existed and it is better advice now.