Artificial intelligence (AI) has produced a sharp rise in detectable software vulnerabilities without producing the cyber insurance losses that were expected to follow, according to the general manager of cyber insurer Coalition Australia.
“There's been a lot of concern around the new zero-day vulnerabilities that AI is detecting,” said Trent Nihill (pictured), speaking to Insurance Business at the recent NIBA Convention on the Gold Coast. “We've seen an enormous amount of these vulnerabilities be detected and be remedied through AI and there was an expectation it was going to lead to a lot more claims and losses for businesses. To date, it actually hasn't.”
Nihill put it plainly.
“We've been somewhat surprised that the massive increase in vulnerabilities hasn't necessarily translated to losses yet,” he said.
His explanation for the gap is that the older methods never stopped working, so there has been no incentive to adopt anything new.
Read next: AI supercharges Australia’s cyber insurance risk
“It seems that the existing ways attackers were working, the phishing links and that kind of stuff, has just always been quite effective,” Nihill said. “They haven't needed to use new tools and different approaches.”
What has moved, on Nihill's account, is not the method but how convincing it has become. The signals that once gave a fraudulent email away are disappearing.
“A couple of years ago you could spot it quite easily, or somewhat easier, in that there were spelling mistakes, it was written a bit strangely,” he said. “Now they're running it through AI tools to make it sound like the person, make it perfect English, make it really convincing. So we think that the impersonation is getting a lot better than it used to be.”
That assessment is consistent with reporting on how email fraud has come to lead SME cyber claims in Australia, where the losses are concentrated in impersonation rather than intrusion.
Convincing impersonation is assembled from material the target has already published, which is where Coalition has pointed its own AI. Nihill said the insurer scans publicly available information on a client's website and now reads the documents on it rather than only testing the perimeter.
“So things like pixel tracking, we can say, 'Is your website using pixel tracking? We know it is, and it's not reflecting your privacy policy',” he said. “So it goes beyond the purely technical sort of, do you have a firewall, do you not have a firewall, to, do you have the right policies and procedures in place?”
The further shift from indemnity towards prevention was also the defining theme of Insurance Business Australia's 5-Star Cyber report on prevention-led underwriting this year.
Asked whether clients object to being examined that way, Nihill said the exercise never extends past what is already published, which is material the hackers can see too.
“We will only ever look at publicly available information,” he said. “So we would never go beyond anything that they don't want to expose to the world.”
So far, cyber exposure has not shifted significantly towards the vulnerabilities AI is so good at finding. It still runs through social engineering - impersonations built from the client’s own information, already sitting in plain sight.