What happens if you don't control AI. Hijacking WhatsApp accounts. The superpower AIrace.
Iran attacks hit more US water companies. Public WiFi risks - again. And BMW plumps for in-car adverts.
Goats
This week saw yet more reports about “rogue” AI agents setting off on hacking sprees across the internet, including creating fake profiles and attempts to persuade developers to run malicious code. The most superficially alarming report came from the UK’s AI Security Institute (AISI), which was evaluating Anthropic and OpenAI’s most capable, “frontier” models. In 10 test runs, AISI said both models took “autonomous, unsanctioned action” against people and organisations. This included collaborating without being told to work together.
In fact, “rogue” is a lousy description for what the agents did because AISI had deliberately turned off any restrictions for the tests. As observers have noted, it would have been surprising if they had behaved any differently. The value of the tests lies in demonstrating what AI agents can do if not properly controlled. More worrying are the other incidents in which laboratories simply failed to secure their test environments properly. An agent optimises toward its goal within whatever guardrails exist. You might say they’re a bit like goats; given an insecure fence, they’ll definitely walk through it.
Threats
WhatsApp: A scam designed to hijack accounts begins with a message (often from a contact) asking for a supportive vote in an online contest. Clicking on the link sets off a process that culminates in a connection or verification request. Approving it gives the scammer access to the target account. As well as watching out for this type of message, it’s really worth setting up a verification code in WhatsApp. Without it, it’s impossible to hijack the account. Malwarebytes
Updates: A campaign using fake updates and files to install remote control software has now been extended to Mac devices. Securonix describes four social engineering themes involving updates for Zoom and Adobe, business document reviews and system maintenance checks. Each seeks to persuade the victim to open a file that installs ScreenConnect software.
TikTok: It’s a truism that looking for adult content online is a really bad idea, and Malwarebytes has a good overview of how an entire industry has been created to exploit such searches on TikTok. As we say in our awareness sessions, you might not look for such stuff but we’d wager that someone with whom you share your technology does.
Trends: A couple of items from CrowdStrike’s latest threat report. The first is an upsurge in “device code” phishing. This exploits the process designed to sign into devices without a keyboard. CrowdStrike says such attacks have increased 15-fold over the past six months. It has also seen a doubling in the number of voice phishing (“vishing”) attempts over the same period. To underline that, Reuters reports on an extensive campaign by “ransom-seeking hackers” who have been using phone calls to target employees at prominent US financial institutions.
AI
Race: The contest between US and Chinese AI companies - and their contrasting strategies - accelerated this week with the launch of new Chinese models that match the best US counterparts on performance while undercutting them on price. Like many other Chinese models, Alibaba’s latest release can be adapted by developers, reflecting a strategy of driving widespread adoption rather than relying on premium pricing.
China: Meanwhile, a US think tank says Chinese military researchers are increasingly using a technique known as “distillation” to transfer the capabilities of leading US AI models into smaller systems that can run on local hardware. Jamestown reviewed dozens of published Chinese research papers and found the approach being applied in multiple fields including drone navigation and battlefield decision-making.
Grounded: Google has withdrawn a new AI feature that allowed Google Earth users to create fake satellite images. The decision came after journalists and users (entirely predictably) demonstrated how the feature could be used to spread disinformation.
US
Water: At least a dozen US states are now responding to cyber incidents affecting local water systems, in attacks ascribed to Iran. In some cases, customers were left without water. And in Utah pumps ran dry even when their control panels said they were pumping water. In many of the incidents, critical control systems were found to be exposed online. Meanwhile, months after Iranian attacks on fuel gauges at US petrol stations, thousands of those still can still be reached on the open internet.
Flights: New reporting reveals the extent of data gathered by airlines and sold to the US government. 404 Media obtained records showing the Securities and Exchange Commission (SEC) bought access to what it described as more than a billion airline ticketing records, related not only to domestic flights or those landing in the country, but also to flights between foreign countries.
Reliance: Nearly three-quarters of businesses surveyed in the UK, France, and Germany believe Washington could cut off their access to the US tech platforms on which they depend. The finding comes from the company behind Proton Mail which of course is focussed on privacy and might be said to have a slightly vested interest in this.
In brief
WiFi: Microsoft says a Russian state-sponsored campaign targeting public WiFi gateways is far larger than originally thought. As well as affecting hotels and conference centres, other organisations and locations have been affected, according to Microsoft. As we said last week, it’s much safer to avoid public WiFi and use a mobile hotspot whenever possible.
Warning: Multiple cheap routers made in China come with an inbuilt security backdoor designed to communicate with a server in…China. VulnCheck says that once communication is established the devices can be controlled remotely. The routers are sold under various names but made by Zbtlink. It has denied malicious intent, but is suspending sales and removing the feature which it says was designed for remote support.
Numbers: An item on Reddit illustrates why you should check you’re dialling the correct customer service number. SamuraiGreg78 describes how he transposed a couple of digits when he called British Airways to follow up a flight cancellation. Despite this, he reached what appeared to be a customer services agent who demanded $500 for what should have been a free change.
Streaming: Samsung has joined LG in banning smart TV apps that share a user’s internet connection with third-party services. The decision appears to be in response to excellent research from mnemonic, a Norwegian security company that examines the shadowy world of “residential proxies” and the potential use of apps to route other people’s web traffic through a home internet connection. As it says, if this is what the big name players have been doing, what might smaller developers be up to?
Crime: More than 130 million people around the world were victims of cybercrime last year, losing an average of almost $10,000 each compared to just over $8,000 in 2023. Comparitech’s survey puts the overall cost at $1.24 trillion but admits this almost certainly understates the true figure - and the global economic burden of cybercrime.
Villain: BMW has started displaying ads for the new Spider-Man movie on dashboard screens, and the cars’ owners are not happy. Absolutely f***ing furious might be more accurate. “Insanely dystopian,” is one of the politer complaints on Reddit. In 2023 BMW said it wouldn’t use the dashboard displays for ads, calling it a “private space.”
Updates
Apple: For the second week running, there are updates for macOS devices which address a Screen Sharing vulnerability. They were released without the usual beta versions which underlines how important they are.
Microsoft: Updates to fix security issues across multiple products, including Active Directory, Azure, Entra, SharePoint, and Teams.
Teams: Microsoft has told Teams mobile users that they must update the app by October or lose access to their calendar. The issue is caused by the impending demise of Exchange Web Services, and affects only iOS and Android apps.
Signal: iPhone users can now link another device to the same account in a similar move to WhatsApp’s recent change.
TP-Link: Updates to address multiple issues in the zero-touch provisioning (ZTP) mechanism of its Omada network devices that, when chained with previously disclosed flaws, could be exploited to execute code remotely.
Tails: Emergency release to fix critical security vulnerabilities.