VPNs, 'old man' masks, and AI: The holes in the social media ban and their fixes

Leading facial age estimation tools were easily fooled by a $22 "old man" mask, a Guy Fawkes mask, and other cheap party costumes, researchers have found.

From December 10, social media platforms will be forced to block under 16s in Australia from holding accounts, but experts warn age checks can be bypassed, using VPNs, AI, and a range of basic disguises.

Facial age estimation tools scan and analyse a person's face in order to guess their age, and is one of several technologies the ban will depend on.

"Every age assurance vendor that we tested had one bypass that was easily accomplished with things that you could buy at your local $2 shop," said Professor Shaanan Cohney from the University of Melbourne.

The preliminary findings have emerged from the early stages of a study being run by the University of Melbourne and Princeton.

The study has been testing three top-performing facial age estimation systems which are seen as likely to be used in the Australian ban.

Participants used tactics such as pointing the camera at video game characters, pulling silly facial expressions, as well as cheap disguises.

They also tested somewhat more realistic "old man" masks, which were bought online for between $20 and $50.

Most of the tactics did not work reliably, Professor Cohney said, but every face scanning tool they tested could be tricked with repeated attempts.

"I got it to work a bunch of times, the students [participating in the trial] got it to work a bunch of times, and this was over 20 minutes of playing around," he said.

"So, assuming that you get to keep trying to prove that you're over a given age, you can probably get away with it."

He said the research team had considered the risk they may end up providing an instruction manual for getting around the ban.

"But because these techniques are so rudimentary, these are things that people are going to be trying anyway," he said.

"All we're going to do is actually provide a fair-minded and transparent assessment of how well the techniques actually work."

The government says it's confident age assurance technology is robust and effective, citing the results of an extensive trial, which were released in full in September.

But Professor Cohney said the government's results "lack the transparency to actually know if what they're claiming holds any water".

"The first thing I would note is [the government's age assurance technology trial] was an industry-run study," he said.

"And just like every banking study shows you that the banks are secure and yet we see banking breaches and bank failures, we'd expect to see the same thing with [age assurance] providers".

The trial was led by a UK-based industry accreditation body, called Age Check Certification Scheme, which describes itself as an independent third party organisation.

The global industry body, the Age Verification Providers Association (AVPA), said facial age estimation providers all have defences in place for such attacks.

"The certification scheme has a lab, where it tests for these things, and they've got all kinds of crazy masks," said Iain Corby, the executive director.

He said no system would stop every bypass attempt, but the success of such novel hacks is often overstated.

"Some of [the providers] might spot 97 per cent of attacks, others might spot 98 per cent of the attacks," he said.

Mr Corby said while a $22 costume mask is "not a very sophisticated attack", he would be surprised if it worked on the most up-to-date models.

"Sometimes [researchers are] using demo systems which don't have all the security controls switched on," he said.

Professor Cohney said his team used the testing interfaces provided by age assurance companies to demonstrate their technology was "bug free and vulnerability free".

The University of Melbourne study is ongoing and Professor Cohney emphasised that the findings about face-scanning tools were only preliminary.

"We're not going to make any strong claims about that until the data is actually in," he said.

Facial-age estimation isn't the only tech that will be used to determine the age of Australian social media users.

Other options include providing official ID such as a drivers license, and "age inference", where platforms can guess a person's age using existing data, such as their online behaviour.

"The fact that social media platforms can use a pick-and-mix of these different technologies, it gives us confidence," said the Communications Minister Annika Wells last month.

However face scanning tech will likely play a bigger role for Australians under the age of 20.

Many won't be able to provide formal ID, others will choose not to, and age estimation may not always be accurate enough.

"I think it's likely that [age assurance technology] will actually prevent a reasonable proportion of bypasses," said Professor Cohney.

"However, for anyone with a modicum of willpower and intelligence, there are going to be methods to get around it."

'That's been fixed': The UK's viral video game hacks

Australia's ban is often described as a world first, but the UK passed similar online age check laws in July, and several hacks went viral within days.

The UK's scheme forces websites hosting adult content, including platforms such as X, Reddit, Discord and dating apps, to conduct "robust" 18+ checks on all internet users.

From day one, social media was awash with tips for bypassing the technology, such as using IDs discovered online or tweaked with AI.

Reddit user Chrisdh79 said he had submitted an image he found by searching 'UK driving license example', and it was "accepted with no issues".

"According to Reddit, I am now Mr Mozahid from London, born in Oct 1989," he said.

A Reddit user posting under the name AncientProduce said he used a drivers license example "with a bit of AI manipulation" to pass the check.

Falsifying identity documents or using another person's ID can be unlawful in Australia.

Many reported cheating age checks to access Discord with an image of a character named Sam Porter Bridges from the video game Death Stranding 2, using a setting called photo mode to manipulate the avatar's face.

"I used Johnny Silverhand to pass the selfie scans," said another user, referencing a character from the role-playing game Cyberpunk 2077.

The ability to move an avatar's face in specific ways, such as opening and closing the mouth on demand, allowed people to bypass "liveness" tests, designed to prevent the use of static images.

Others reported being unfairly logged out, despite using the system in good faith.

"I'm literally 26 and the ID verifications didn't work," said Reddit user Kyraapd, adding that he'd ultimately given up and instead gone to a porn site with no age check in place.

The Age Verification Providers Association (AVPA) said the industry quickly discovered solutions for the best-known hacks, including gameing avatar bypasses.

"That's been fixed, as soon as we noticed that was a problem," said Iain Corby from the AVPA.

He says it's possible not every provider had patched those weaknesses yet, but that "everybody is now in a position to [make] improvements in their defenses".

Professor Cohney said there were early indications in his research that game-avatar bypasses were still intermittently effective.

"Right now, even video game characters which are not that photorealistic can trick a lot of the systems," he said.

He warned those attacks would become harder to beat as AI continued to improve.

"The ability to produce avatars or AI generated people that can do that ... is going to rapidly improve over the next number of years," he said.

"It's not that every 12-year-old is going to be writing AI code to put a filter on their face.

"All you need is one person, somewhere in the world to have written a tool to do it, and then everyone will be able to use it."

Mr Corby laid much of the blame for faults in the UK rollout with social media platforms, accusing them of deliberately weakening age checks in some cases.

"So they allowed, for example, people to have as many attempts as possible at trying to pass a facial estimation," he said.

"And that allowed the hackers to effectively test and find different ways around it."

In another case he said a "client had requested that a lot of the security controls were removed".

"So obviously, if you turn off the tech, it's not going to work," he said.

Mr Corby said Australia's scheme had a better chance of success on that front, because the eSafety Commissioner Julie Inman Grant has powers to check platforms are doing the right thing.

"It may be less likely that they take shortcuts," he said.

However unlike the UK's scheme, Australia's does not require providers to be accredited, and eSafety has not set any minimum standards for accuracy.

Mr Corby is hoping that will change when the ban is reviewed next year.

"Australia has second-mover advantage … you can learn the lessons from elsewhere, " said Mr Corby.

"So hopefully we will have a less traumatic implementation [than the UK] come December the 10th".

Who's winning the cat and mouse game on VPNs?

Another well-known vulnerability in the ban is VPN technology. It allows a person to disguise which country they're in by issuing them a new, temporary IP address.

Most VPNs are easy to use and some are free, although experts warn that free and low-cost VPN providers usually harvest and sell their users' data.

VPN downloads surged after the UK's age check laws were implemented, with some companies advertising their products as a way to "beat the UK porn ban".

In the weeks that followed, there were an estimated 3 million new VPN installations across the UK, according to a report from the AVPA.

The report noted there was no evidence that the VPNs were being used by children, but said there was a long-term risk that VPN use "may undermine the efficacy of the [UK's Online Safety] Act".

One in three Australians already use VPNs and 74 per cent are aware of the technology, according to a survey by one of the major VPN providers, NordVPN.

The company said it's expecting a surge in Australian users, come December.

"Whenever a government announces an increase in surveillance, internet restrictions, or other types of constraints, people turn to privacy tools," said a spokeswoman for NordVPN, Laura Tyrylyte.

The eSafety Commissioner Julie Inman Grant says under the Australian scheme, platforms will be expected to monitor for VPN use and put solutions in place.

"We've been certainly watching what's been happening in the UK, how [VPNs] have surged," she said.

"But this is a totally different scheme."

Guidelines issued by eSafety spell out some of the solutions for social media companies, such as monitoring for IP addresses linked to VPN providers and cross-checking against geolocation data and behavioural patterns.

Professor Shaanan Cohney said detecting VPN use was the easy part, but determining a person's true location or home country was more difficult.

"It's another thing altogether to tell that they are regularly resident in Australia, which is what the legislation actually requires," he said.

Mr Corby was more optimistic that platforms would be able to work it out.

"The great thing about social media is it does have all of this rich information on its users," he said.

Laura Tyrylyte, from NordVPN, said it was theoretically possible, but the company hadn't yet seen social media platforms trying to do it.

She also warned that if platforms get it wrong, there could be global consequences.

"Blocking IP addresses from reputable VPN providers would disproportionately affect millions of legitimate users outside Australia who rely on VPNs for privacy and security," she said.

Professor Cohney said there was no question the technology would be effective in many cases.

"That's not the question," he said.

"The question is: is the technology effective enough to make it worthwhile for all the problems that are going to come along with it?"