Hackers used AI to exploit an unknown flaw in widely used software for a “mass exploitation” event, Google has revealed.
The tech company’s experts say it is the first time they have seen cybercriminals successfully use AI to develop a “zero-day” vulnerability, which refers to a previously unknown bug that requires immediate action to fix.
The hackers found a way to bypass two-factor authentication on a popular piece of software, but were thwarted by Google’s Threat Intelligence Group.
Two-factor authentication is where a user has to confirm their identity using two different types of evidence, usually a password plus a code sent by text or through an app, such as Google’s own Authenticator app.
The software and vendor are not being named, but the flaw has been fixed, Google said.
The discovery heightens fears that AI is helping hackers find and develop flaws in commonly used software. It comes after Anthropic, the developer of a new AI model called Mythos, said the new program was too dangerous to release because of its capabilities.
The UK’s AI Security Institute (AISI) tested Mythos and called it a “step up” in the abilities of AI to carry out an autonomous cyberattack from start to finish.
The hackers were eventually thwarted by Google’s Threat Intelligence GroupAlamy
The institute also revealed that ChatGPT’s latest 5.5 model, which was released in April, “reaches a similar level of performance on our cyber evaluations” as Mythos.
Adam Beaumont, the AISI’s interim director, told the Centre for Emerging Technology and Security conference last month of its testing results: “Two years ago, that finding probably would have sounded quite wild. I think today it sounds sobering. In two more years, it will be routine.”
Google said it did not believe the criminals it discovered were using Mythos or the Google Gemini AI software, but had leveraged an older system.
In its AI threat tracker report, the company detailed how hackers were using the technology to industrialise attacks, from hunting for software flaws to building malicious code and automating parts of an attack that could continue around the clock.
John Hultquist, the chief analyst at Google Threat Intelligence Group, said: “We believe this is the tip of the iceberg. If criminals are doing it, then state actors with significant resources probably are too.
“Each new generation of models will reduce the need for expert-developed harnesses, but they are almost certainly out there. We have to recognise the limits of our visibility into the backend of spies and criminals. The signs won’t be obvious. The race has started already.”
Gavin Millard, the vice-president of product at Tenable, a cybersecurity company, said that AI had reduced the time from the disclosure of a flaw to its exploitation by hackers.
He said: “That disclosure of vulnerability to exposure of a real issue that could be leveraged by attackers has shrunk dramatically. A few years ago, it was, on average, 60 or 90 days. That has completely compressed to no days … it’s minutes or hours.”
However, he added: “We’re not at that point where the attackers are suddenly going to be able to take everything out. I’m not worried about that kind of stuff. I’ve not built a bunker in my garden that isn’t connected to the internet.”
Both Millard and Hultquist were optimistic that AI would eventually be a boon for cyberdefence, scanning software before release for bugs and automating cybersecurity tasks.
Hultquist said: “We actually believe that as AI development becomes more commonplace, we’ll be able to use the state-of- the-art models to basically proofread new software at birth. We’re going to be at least taking a big bite out of one of the major vectors for threat actors. The only issue is that we’ve already put a lot of software out into the world.”