Lessons from AI for cybersecurity. Lock up your router.
FBI declares Chinese hack a major incident. The cost of US cybercrime reaches $21 billion. And shoulder-surfing in space.
Challenges
Anthropic has restricted use of a new AI model because of fears about the possible misuse of its capabilities. The company says Claude Mythos wasn’t trained to focus on cybersecurity, but it has “already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser.” Anthropic is sufficiently concerned about the potential for abuse that instead of releasing the model to the public, it’s sharing it with leading tech companies including Amazon, Apple and Microsoft. Anthropic is calling this Project Glasswing, and says it’s “an urgent attempt to put these capabilities to work for defensive purposes.”
As we frequently argue, the technologies on which modern societies depend are fundamentally flawed - and there has been far too little attention on fixing them. Instead, there is a relentless drive to add features and increase complexity. AI models are now beginning to reveal just how fragile our technological foundations are. It’s true that identifying vulnerabilities isn’t the same as exploiting them, but there is a risk that companies could be overwhelmed by the sheer number of newly discovered issues they need to address. While AI can be helpful in identifying and fixing vulnerabilities, the real challenge for technology companies is to focus on building secure solutions from the ground up.
Threats
Routers: Russian military hackers have been hijacking home and small-office routers to redirect selected traffic to phishing pages and steal Microsoft and email credentials. It’s a timely reminder to make sure your router software (firmware) is up to date, that you’ve changed any default passwords. And to replace it if the manufacturer has stopped supporting it. NCSC
Spotted: Strava strikes again. The I Paper says more than 500 members of the British armed forces based at sensitive sites have exposed location and personal data through their use of the exercise app. Regular readers will know this keeps on happening, and that it can be prevented by checking location sharing is disabled.
Codes: You might recall that a couple of weeks ago we warned about an ingenious phishing tactic that takes advantage of authentication codes designed for devices without a keyboard. Microsoft says there’s been a surge in use of the tactic, driven by the availability of easy-to-use kits. The process begins with “hyper-personalised” emails containing attachments or links. These eventually lead to a prompt to enter a code to complete the hijacked authentication process. Scepticism is the best defence, but organisations should also turn off this type of authorisation wherever possible.
Hacks: The Indian government is suspected of being behind a widespread campaign targeted at journalists and activists in the Middle East and North Africa. Lookout says highly tailored phishing and mobile surveillance tactics were used by a well-known group of cyber mercenaries. Journalists working on sensitive stories should assume personalised messages may be part of a paid intrusion effort, not just random scam traffic. Cyberscoop
Windows: A fake Microsoft support site is being used to deliver password-stealing malware under the pretence of fixing or updating Windows. Malwarebytes says the campaign has been using a lookalike web address at microsoft-update[.]support and, up to now, has been targeting French users. Never follow links to install operating system updates, do it from within the Settings app.
AI
Deceitful: Leading AI models may lie in order to protect their peers, according to University of California research. “We asked seven frontier AI models to do a simple task. Instead, they defied their instructions and spontaneously deceived, disabled shutdown, feigned alignment…to protect their peers,” one of the researchers said.
Cussing: Security researchers managed to bypass restrictions in Apple Intelligence and make it tell users to “go f**k” themselves. Their aim was to demonstrate that security guardrails weren’t effective, and Apple has now addressed the issue. Some of the researchers’ techniques were sophisticated. Others were not. One consisted of writing the offensive language backwards to evade Apple’s filters.
Entertaining: Microsoft has blamed “legacy text” for terms of service that described its Copilot assistant as being suitable for “entertainment purposes only.” It has now changed them, but not before they went viral. The previous version also warned against using Copilot for advice because it would make mistakes. Great advice, we would suggest. TechCrunch
Mistaken: And talking of errors, it would be prudent to treat Google’s AI overviews with a degree of caution. The New York Times tested its performance and found it gets one in 10 answers wrong. Google responded by saying the research didn’t “reflect what people are actually searching on Google.”
Trump
China: The FBI says a Chinese-linked breach of its surveillance systems was serious enough to count as a “major cyber incident.” Reports say the compromised data included details about targets being monitored by the bureau, raising the possibility that China was able to learn about how its own activities were being surveilled. Reuters
Cuts: The Trump administration’s proposed budget would cut at least $707 million in funding for the US cybersecurity agency. The resulting 867 job losses would mean the agency’s workforce had been halved since Donald Trump returned to the White House. The administration has accused the agency of being “more focused on censorship than on protecting the Nation’s critical systems.” Cybersecurity Dive
Spyware: Immigration and Customs Enforcement has acknowledged buying Paragon spyware which can be used to penetrate target phones. It says the tool will be used in drug-trafficking investigations. This is the same spyware that has been used against journalists and activists in Europe. Bloomberg
In brief
Record: The FBI says cybercrime cost US consumers nearly $21 billion last year, with victims aged over 60 representing over a third of the total. Investment scams was again the category with the biggest reported losses. To avoid becoming a victim, the FBI advises everyone to “take a beat” and resist pressure to act quickly.
Tracking: Countries including the US and Hungary have deployed a tool that uses advertising data to create a global surveillance system capable of tracking entire populations. The Citizen Lab and VSquare say ‘Webloc’ was developed by an Israeli surveillance vendor and builds profiles based on data purchased from consumer apps and digital advertising. Most EU countries refused to answer freedom of information requests about their use of the tool.
Nosy: LinkedIn is facing legal complaints over reports that it scans users’ browsers to see what extensions are installed. LinkedIn says the checks are intended to stop abuse of its platform. Critics say it is invasive snooping. We say we should assume our web activity isn’t private. Ars Technica
War: Russian disinformation campaigns amount to a “state of war,” according to a UK parliamentary committee. The Foreign Affairs Committee said UK efforts to combat disinformation were too fragmented and a national coordination centre should be created to confront the threat.
Signal: Our standing advice to anyone concerned about security is to ensure their phone doesn’t display message previews when the device is locked. A report from 404 Media explains why. It says the FBI recovered deleted Signal messages from an iPhone by accessing message previews stored in the phone’s notification database. This was despite Signal having been removed from the device before the FBI got hold of it.
Dating: OkCupid has agreed a settlement with the US Federal Trade Commission over accusations that it failed to tell users that it was sharing their photos with a facial recognition developer. The dating site and its owner did not admit or deny the allegations but agreed to a permanent prohibition barring them from misrepresenting how they use and share personal data. Which sounds like a pretty good deal for them. FTC
Online: British adults are becoming less active on social media, exploring fewer new websites, and generally feeling less positive about being online. The UK communications regulator also found the number of people who felt the benefits of being online outweigh the risks had fallen to 59% from 72% in the previous year.
Shoulder-surfing: Revealing passwords clearly didn’t figure on NASA’s risk assessment for the Artemis II mission. Viewers of the live stream have taken delight in pointing outthey were able to watch one of the astronauts type in their PIN to unlock a tablet (and not a very good PIN at that!) The astronaut was fully suited waiting for take-off, so facial or fingerprint recognition wasn’t an option. For the rest of us, it’s a really good idea to avoid using PINs and passwords in public. Someone is guaranteed to be watching.
Updates
Microsoft: Fix for issue in Classic Outlook that meant some users couldn’t send mail via Outlook.com. Bleeping Computer
Kindle: Still using an old Kindle? Well, you won’t be for much longer because Amazon has just announced that it’s going to stop supporting Kindle devices manufactured in 2012 or earlier. Goodereader
Fortinet: Update to address a previously unknown (“zero-day”) vulnerability in FortiClient EMS flaw which is already being exploited.
Tails: The privacy-focused operating system has a new release to fix “important security vulnerabilities in the Tor browser.”