AI warnings get louder. Invisible phishing. An AI worm. LG's inquisitive smart TVs.
And the perils of public screens.
Doom
Last week we reported on concerns that increasingly capable AI systems are behaving in ways their developers neither intended nor anticipated. This week, Anthropic released a report listing various deliberate attempts to misuse its technology. Among the most eye-catching examples, it says it found criminals and state-sponsored hackers trying to use its Claude models to automate cyber attacks and military drone swarms, conduct mass surveillance, and weaponise a deadly mosquito-borne virus. “As models become increasingly capable, their risks will increase, unless AI developers and society’s defenders act to make them safer,” Anthropic opines (with masterful understatement).
Anthropic’s report will intensify discussionabout how quickly increasingly autonomous AI should be developed and deployed. This month, an Anthropic researcher resigned publicly, accusing the leading labs of “gambling with our lives”; hours later a top safety researcher at Anthropic put the odds of AI eventually killing all humans at more than 10% within the next 10 years. Such numbers are highly speculative, and some have rejected the warnings as sales talk dressed up as doom-mongering. It’s impossible to put a reliable number on the risk that AI represents. There is one point of broad agreement: the technology is improving faster than anyone’s ability to understand or control it.
In practical terms, AI does continue to be a useful tool - but one to be used with caution. From the start, we’ve said, “Don’t trust. Do verify.” Think twice before giving an “agent” standing access to your data, especially to email, financial data, and passwords. Assume scammers now have access to cheap, convincing AI impersonation, so keep confirming who you’re actually dealing with. Professionally, if you’re deploying agents, stick to the basics: least access, off the open internet unless essential, logged and monitored, with a human in the loop for anything irreversible. And don’t rely on AI to check other AI. Incidentally, we use Claude to proofread this newsletter (though a human checks its work). This week, it refused to complete the task, citing cybersecurity safeguards.
Threats
Invisible: Microsoft has spotted a phishing campaign using a trick borrowed from attacks on AI systems. It uses invisible “Unicode” characters to split words such as “funding”, making them look normal to recipients while obfuscating them from email security systems. While ingenious, Microsoft says other signals enabled it to detect the messages - two million of them a day at their peak.
Blob: Phishing sites are getting harder to spot, because they may not actually exist on the web. Barracuda says attackers are routing victims through legitimate Microsoft services before creating the fake login page locally inside their browser. The technique uses a temporary or “blob” URL (which stands for Binary Large Object, in case you were wondering). The attacks begin with convincing DocuSign-themed emails and even include calendar invitations to add legitimacy. The safest response to an unexpected login or document request remains simple: don’t use the link in the message. Open the service independently and check the request there.
ScreenConnect: Huntress has another example of legitimate remote support software being turned into something resembling a computer worm. It says social engineering (including posing as IT support) is being used to persuade victims to use Windows Quick Assist and install modified versions of ScreenConnect. Once established, the malicious software can automatically transfer and execute its payload on other ScreenConnect systems it connects to. An update has addressed the underlying vulnerability, but the need for caution over remote access software remains.
Students: UK students are receiving texts, emails and calls claiming that the bank details on their Student Finance account have changed. Fake login pages capture genuine credentials; criminals then enter the real account and replace the bank details before maintenance payments arrive. Any message referencing changes to financial transactions should be treated with caution.
Microsoft: Microsoft 365 users are facing an onslaught of attacks, including fake IT support calls, Teams messages and increasingly sophisticated phishing sites. One phishing-as-a-service platform acts as a middleman between victims and Microsoft, capturing passwords, authentication codes and session cookies in real time, allowing attackers to bypass conventional multi-factor authentication. CloudSEK says the campaign has already compromised credentials at more than 250 organisations across 40 countries. The best defence is phishing-resistant authentication such as passkeys or physical security keys.
AI
Worm: Researchers have demonstrated just how dramatically AI could accelerate the development of sophisticated cyberattacks. Security firm Calif used AI to discover a vulnerability in the WeChat messaging platform. In just two days, it used the issue to create a working exploit to run code remotely. A week later, it had built a zero-click worm capable of jumping between iPhones and Android devices simply by making WeChat calls - victims didn’t even need to answer. Calif says work that would previously have taken a larger team months can now largely be done by AI. The issue in WeChat has now been fixed.
China: US security and intelligence agencies have accused China-based AI companies of systematically stealing proprietary information from their US rivals. The NSA, FBI and cybersecurity agency said industrial-scale knowledge distillation campaigns formed the core, not merely a supplement of China’s AI development strategy.
Shadow: The UK’s cybersecurity agency is calling on organisations to tighten up security procedures and crack down on the use of unapproved AI tools. Shadow IT is a long-running issue. Shadow AI has the potential to be a far more serious problem. The agency cites a recent study in which 71% of employees reported using AI tools that their employer hadn’t approved.
Effects: Anthropic has created an interactive tool that lets users test their assumptions about how productive - or disruptive - the AI boom might be. There are a range of scenarios, stretching from a gentle boost to the economy with little effect on workers to soaring GDP and 14% of workers losing their jobs.
Glossary: In case it’s of use, TechCrunch has a guide explaining the terms that are bandied around about AI. It’s a handy reference to the language of AI and terms like “Opaque Recurrence,” “Weights,” and “Transfer Learning.”
In brief
Telly: LG smart TVs are likely to be doing far more than showing pictures. A video from Gamers Nexus found them constantly scanning their local WiFi networks and recording nearby audio conversations even when the TV is set in stand-by mode. The researchers say information collected while offline can subsequently be transmitted when the TV reconnects to the internet. LG is mighty upset about the audio claims, saying its TVs don’t record ambient conversations, though some of the other findings go unanswered. It is clear that LG, in common with other TV manufacturers, go to great lengths to harvest data about their users. You can watch the video and decide for yourself just how far they go.
Extreme: This is how the United Arab Emirates and Israel are said to go about securing communications between their leaders. The UAE intelligence attaché in Israel acquires a new phone and installs a “special” app on it. The phone is brought to the Israeli Prime Minister’s office and, once the leaders have finished talking, the device is destroyed. Which is all very well, but it didn’t stop an Israeli newspaper from reporting that the call contained a warning from the UAE about a major Hamas operation days before the October 2023 attack. Israel has rejected the report as an “absolute lie.” Either way, a reminder that absolute security is a fiction.
Breach: A brief update to last week’s item about the sale of 153 million driving licence scans. Rather than outright admitting it was responsible, identity verification company IDScan has issued a subjunctive-laden statement saying “an unauthorised third party may have accessed and/or copied certain customer information.” And because full access to the information required payment, it’s only “out of an abundance of caution” that IDScan is offering free credit monitoring to the undisclosed number of people affected.
Voice: Dutch police have published a recording of the voice phishing attack that allowed hackers to breach a Dutch ISP this year. Unsurprisingly, the recording is in Dutch, but you don’t need to speak the language to appreciate just how convincing it is. Conversational, natural, pretty much exactly what you’d expect an IT support call to sound like. We all like to think we wouldn’t fall for this sort of scam. The recording is a salutary warning.
Travel: The risk of oversharing location information on fitness apps is well documented (and often covered here). The latest example comes from travel app Polarsteps, from which Follow The Money found it was possible to retrieve roughly 230 million photos/videos and one billion GPS points associated with almost two million journeys. Precise photo coordinates could reveal places such as users’ homes and enabled military personnel to be tracked. As Polarsteps says, it makes clear that if a user selects “Anyone” in privacy settings, that is literally what it means: anyone can see the user’s data, even if they don’t use the app.
Quantum: The G7 group of nations has urged governments and companies to start preparing their systems for a future in which quantum computers could break widely used forms of encryption. While the timeline is uncertain, the group says the “quantum threat” should be reframed from a distant future problem to a near-term threat that affects all sectors, not just critical infrastructure.
Public: Like us, you’ve most likely seen people working on sensitive subjects in public. In the case of the announcer on US gameshow “Wheel of Fortune,” a fellow passenger photographed him viewing a chatroom linked to paedophilia while on a US domestic flight. The Daily Mail got hold of the images. He’s now been fired, despite saying he visited the site by mistake. Of course, there are some sites that shouldn’t be visited at all, but it’s a reminder that people can’t help looking at what other people are doing. If it’s absolutely essential to work on something confidential in public, then at least use a privacy screen.
Updates
Microsoft: Another record set of security patches, addressing more than 960 issues across its products - including two previously unknown “zero-day” vulnerabilities which are being actively exploited.
Apple: New operating system versions for iOS (26.6.2) and iPad (26.6.2). They come just days before the release of Apple’s next major software update, taking the version number to 27. These changes can have unpredictable effects and we normally advise waiting a day or so before taking the update. It’s also good practice to make sure you have a reliable recent backup of the contents of your devices….just in case.
Chrome: Google has confirmed that, like its peers, its web browser will now be updated every fortnight instead of every four weeks. “With automated AI discovery tools and community bug reports generating higher patch volume, shorter release cycles make managing security fixes significantly simpler,” it says. We say, shut your browser down at the end of every day - that way you’ll be sure it’s up to date.
MikroTik: Essential update to address a particularly nasty set of issues that can be used to obtain full administrator access to vulnerable routers. without valid authentication. The main risk is to devices with their SSH remote-access service exposed to the internet.
Patch Tuesday: Monthly security updates fromAdobe,Ubuntu, Cisco,SAP,IBM,HPE,Dell,Ivanti,ASUS,TP-Link, and Fortinet