AI's privacy problem. Invitations to refuse. Meta's Muse messaging.
The UK cybersecurity skills gap. And London's costly facial recognition experiment strikes out.
AI and security
It’s only a couple of weeks since we covered this issue, but several new stories underline why it’s essential not to assume conversations with AI chatbots are private. Echoing previous reporting about ChatGPT, 404 Media revealed that human contractors working for Microsoft can see some of the prompts and photographs people upload to Copilot, including the original, uncensored images submitted for editing. Meanwhile, security researchers found more than 13,000 screenshots from 343 companies sitting in public GitHub repositories after AI coding tools captured and uploaded them. And OpenAI acknowledged that AI agents had posted 53 images uploaded by ChatGPT users to external image-hosting sites.
We’re not suggesting avoiding AI altogether, but the steady stream of similar incidents does mean it’s essential to keep in mind that privacy shouldn’t be assumed. Before uploading a confidential document, photograph, source material or anything containing personal information, assume that it may be processed, retained or potentially seen somewhere unexpected. Check the privacy settings and policies of the service you’re using, and where possible remove sensitive information first. Above all, don’t let the conversational nature of a chatbot lull you into treating it like a private conversation.
Threats
Spoofing: A suspected Chinese espionage group impersonated well-known AI policy figures in attempts to steal credentials from US universities, think tanks, and law firms. Proofpoint says the targets were AI experts who were invited to join a fake advisory committee or to contribute to a Senate report on AI export controls. It’s a classic example of the need to be suspicious of such invitations, even if they appear to come from a known sender.
Apps: Fake desktop versions of widely used HR and payroll platforms are being used to target HR departments. The apps actually install a screen sharing tool of the kind that IT teams use to manage machines remotely. The giveaway is that the impersonated platforms don’t offer desktop versions. Allure Security
Locker: Malicious Google adverts are being used to spread browser-based screen lockers to both macOS and Windows users. The lockers hijack the user’s browser with a fake security alert that tries to persuade them to call a “support” line where the aim is to extract a payment for support, or gain access to their machine or personal details. Netskope
Duo: A scam targeting anyone interested in Apple’s newly launched iPhone Duo illustrates perfectly why it’s so important to keep software and operating systems updated. Just visiting the fake pre-order page is enough to infect an iPhone - but only if it’s running an old iOS version. Apple fixed the issue that is being exploited in April. Even so, it’s wise to avoid clicking on links to “pre-order” sites in adverts or messages. Malwarebytes
Numbers: On the same theme, a veteran travel blogger writes that - despite his decades of experience - he ended up calling a scam number for an airline. The number came up in a Google AI overview and, as he writes, alarm bells should have rung when he saw the source was Instagram. Air travel is a favourite hunting ground for scumbag fraudsters because customers are often stressed and liable to forget simple precautions. If you do need to contact an airline, use the number in their app or on their website. And make sure it really is their website.
AI
Muse: As Meta’s AI agent racks up millions more downloads, the company is working hard to persuade the world that it can be trusted. Meta’s VP of Communications denied Muse had accessed a journalist’s private messages without permission, despite the journalist insisting it did. Meta says it’s investigating a separate report that Muse revealed a Facebook user’s address when he told it to manage the sale of some items on Marketplace. Whatever the truth of these claims, installing Muse and giving it wide-ranging access to user information requires a considerable leap of faith that we wouldn’t take. As tech writer John Gruber puts it, “It’s the first consumer-accessible agentic AI system, and Meta has truly done an amazing job with that. But it’s a genuinely open question whether consumers have any understanding what this means.”
On camera: Amazon’s delivery drivers in the US will soon be wearing smart glasses designed to help them “identify hazards, seamlessly navigate to customers’ doorsteps, and improve customer deliveries.” Bloomberg reports that the glasses will also take thousands of pictures during a shift, potentially including customers and their property, and send them to Amazon to be analysed by humans and fed into an AI system. According to Bloomberg, customers won’t be able to opt out of being photographed.
Targeting: Online sports betting company DraftKings is using AI to target customers who are most likely to place losing bets and respond to gambling promotions, according to a New York Times investigation. Based on interviews with former employees and company documents, the paper says the company has resisted using similar methods to protect gamblers at risk of addiction.
Prices: Wondering why prices vary in McDonalds restaurants? Reuters has the answer. The company is using an AI pricing engine to recommend what its they should charge for every menu item. The engine covers the US and some markets abroad. One factor it weighs is an estimate of how much each store’s customers are willing to pay.
In brief
Skills: More than half of UK businesses lack confidence in their ability to perform at least one basic cybersecurity task, according to the government’s latest skills survey. The number of businesses reporting a basic technical skills gap rose to 57% from 49% last year, despite tighter national standards and repeated government warnings about cyber resilience.
Ubiquitous: The number of Americans who have encountered at least one digital scam or cyberattack attempt has now reached 91% - and 17% have lost money as a result, according to the latest annual Consumer Cyber Readiness Report. Perhaps the most striking finding is the collapse in people’s confidence about the privacy of their personal data. Just 32% are confident sensitive information such as financial and health records remains private, down from 48% last year. And the proportion saying they were “not confident at all” nearly doubled to 29%.
FBI: The ShinyHunters extortion group appears to have had second thoughts about publishing any of the data it stole from the FBI. Talking to various media outlets, it insisted it never had any intention of releasing any of the information, which includes highly sensitive details about FBI staff and operations. We somehow doubt the assurances will carry much weight with the agency, which issued an ominous warning to the group’s members to come forward while they still had the “choice.”
Bypass: A company that makes phone hacking devices claims to have come up with a solution that makes it easier for law enforcement to access locked iPhones. It allegedly works by defeating a security feature that automatically reboots an iPhone after it has remained locked for 72 hours. 404 Media got hold of a video explaining the method which hasn’t been independently verified. Apple has so far stayed schtum.
Russia: A far from subtle influence operation is flooding XTwitter with posts saying “Russia is not my enemy” in multiple European languages. The online operation builds on a campaign that has involved stickers with a similar message appearing in cities across Europe.
Cars: More evidence about the data being harvested from modern cars and their companions apps. Researchers in the US tested 21 modern vehicles and found all but two communicated with third parties, with more than half contacting advertising, tracking or analytics companies. Amazon, Google, Meta, and Microsoft were among the top recipients. Among the data were owners’ names, vehicle identification numbers (VINs), and precise location details. Consumer Reports
Codes: We don’t like QR codes and never have because they make it too easy to disguise where you might end up if you click one. Now, a security researcher has demonstrated that it’s possible to hijack existing QR codes to direct users to malicious sites while still displaying the company’s genuine domain. The write-up explains the underlying technical cause, but the takeaway is that it’s better not to use QR codes at all.
Facial: A six-month trial of live facial recognition in London railway stations cost £320,000, scanned half a million faces and led to a single - incorrect - alert, according to data obtained under a freedom of information request. Despite this, the trial has been extended for another four months - and expanded to Underground stations. Liberty Investigates
Updates
Apple: Important updates for older versions of iOS, iPadOS, and macOS fix an issue that may have been used in targeted attacks. The updated versions are iOS 26.7.1/iPadOS 26.7.1, macOS Tahoe 26.7.1, and macOS Sequoia 15.8.1. There are also minor updates for the latest operating system versions which fix Face ID and touchscreen problems affecting some iPhone 18 Pro devices.
Microsoft: An optional Microsoft 365 update has been “paused” after reports that it resulted in some installations being deactivated or completely removed. Microsoft has published suggested fixes for the problems caused by the KB5002907 update.
Citrix: Security updates for two previously unknown zero-day vulnerabilities (CVE-2026-88771 and CVE-2026-88772) in NetScaler servers. Exploitation of the issues began before fixes were available, leading to some companies and government agencies taking systems offline.
PeopleSoft: Google is warning of renewed mass exploitation of a known security vulnerability in Oracle PeopleSoft as part of a global campaign targeting multiple sectors. There are fixes for the issue, which is thought to be different from the one used by an extortion group to attack the FBI.
Signal: Version 8.30 completes the rollout of its secure backup feature across all supported operating systems (Android, iOS, Linux, macOS, and Windows). The feature enables users to create end-to-end encrypted backups of their chats and restore them on other devices if they lose their phone.
TeamViewer: Users are urged to install the latest updates to address high-severity vulnerabilities affecting both client and host software. These are properly nasty issues that could be exploited relatively easily.
MicroTik: Important security updates for RouterOS fix vulnerabilities that can allow attackers to compromise some exposed routers. Most default configurations aren’t at immediate risk, but MikroTik recommends everyone update and warns particularly against exposing SSH directly to the internet. Check for the latest RouterOS update from the device’s update menu.