Meta makes AI personal. The $1.6 billion US law enforcement scam.
OpenAI hacks Australia government websites. FBI investigates database breach. And being rude to cyber crooks.
Personal
Just over two weeks after launch, Meta’s Muse AI agent remains the most popular free iPhone app in the US (ahead of ChatGPT), making it the first personal agent to break through to a mass consumer audience. Unlike a conventional chatbot, once connected to services like email, calendars, social media and shopping, Muse is designed to act on the user’s behalf. Or, as Meta puts it, “Muse works in the background to move your life forward.” All this obviously requires intimate access to someone’s digital activities. Meta insists Muse conversations and private data aren’t fed into its advertising systems, and users can opt out of having interactions used to train its AI. Despite that, as The Verge discovered, once connected to Instagram, Muse inferred details about the reporter’s interests and identified her location from an Amazon order.
Needless to say, Muse is not a philanthropic exercise.Meta shares are up more than 20% since the launch on September 8, adding more than $300 billion to the company’s market value. Financial analysts say Muse could bring “substantial new monetisation opportunities for Meta in the form of advertising, subscriptions and transaction revenue-share.” (Amazon won’t be contributing to this because it has banned Muse from making purchases, saying Meta didn’t seek permission.) While there are already highly capable AI agents, Muse is the first opportunity for most non-technical users to experience a deeper integration between AI and their everyday lives. It’s an integration we’re happy to resist for the moment, particularly in the form offered by Meta’s Muse.
Threats
Fakes: Scumbags impersonating law enforcement or government officials have cost victims more than $1.6 billion since January 2025, according to the FBI. Its Internet Crime Complaint Center (IC3) received nearly 61,000 complaints of this type between January 2025 and July 2026, putting the average loss per complaint at more than $26,000. Surprisingly, the most common type of scam involves criminals managing to convince victims to pay for the removal of non-existent criminal charges.
Gift cards: Bitrefill sells gift cards for popular stores like Amazon, Deliveroo, Apple, Nintendo, and thousands of others. (They also sell eSIMs and mobile top-ups.) It’s currently being targeted by scammers who have created lookalike websites that take victims through what appears to be a normal purchase but of course simply sends funds to the fraudsters. Malwarebytes
Calendar: There’s been a remarkable rise in the use of calendar invites as a phishing method. Security firm Sublime says there’s been a steady increase over the past few months but September is on course to be 28 times higher than August. It’s a particularly effective trick because malicious invitations can sometimes reach your calendar even when the accompanying email is blocked. Treat unexpected calendar invitations, and especially links or attachments inside them, with the same suspicion as unexpected emails.
Sharing: You may have come across apps that pay people to share their unused internet connection. They make for an attractive side hustle but they could also create a serious security risk, especially in a work environment. Silent Push found one such service might expose systems normally hidden from the internet. The apps are legitimate software, so antivirus may not flag them.
YouTube: Scammers have come up with an ingenious way to make people drain their own crypto wallets without sending them a phishing link. It involves professional-looking YouTube tutorials showing people how to use Claude AI to build an AI crypto trading bot. Anyone who followed the tutorial and compiled the code saw their funds disappear. TRM Labs
Claude: Meanwhile, a convincing offer of a free Claude Max subscription is being used to steal Google accounts. The phishing site displays what appears to be a genuine Google login window, complete with the correct address and padlock, but the entire window is fake and drawn inside the webpage. It’s vital to check the URL in the browser’s real address bar, and a password manager also defends against this type of attack because it won’t fill your Google password on a fake site. It’s also worth noting that you can drag a real sign-in pop-up outside the browser window, but a drawn fake one stops dead at the page edge.Malwarebytes
AI
Australia: You’ve probably seen stories about OpenAI’s unauthorised access to Australia’s national healthcare statistics portal. Such events are no longer surprising, although this appears to be the first instance of an AI agent autonomously hacking a government system. What is remarkable is that it took OpenAI almost three months to tell anyone in Australia what had happened. And when it did, it sent an email to a generic address rather than inform anyone in particular. This suggests not only a failure to control its technology, but also a serious failure of incident response. Not to mention the possibility that there are other incidents that we don’t yet know about.
Hallucination: The US narrowly avoided boarding a Chinese ship based on “entirely false” US intelligence generated with the help of AI tools. CNN cites four sources as saying US forces were preparing to intercept and board the ship before officials discovered the error. CNN says a chatbot had combined open-source intelligence with secret signals intelligence to inaccurately identify the material the ship was carrying as nuclear arms program components.
Whistleblowing: SecureDrop, which provides anonymous communication systems for journalists and sources, is now recommending that news organisations explicitly warn potential sources not to use AI chatbots to research, draft or prepare material connected with whistleblowing. The same warning applies to logged-in search engines and other services tied to a person’s identity. The risk stems from the digital trail created by chatbot conversations, searches and other activity linked to an account: this could potentially help identify a source, even if the eventual submission is anonymous. Its advice is to use Tor Browser for all whistleblowing-related activity, rather than relying on SecureDrop as the sole means of identity protection.
Warning: AI agents could be useful to defenders, but the UK’s cybersecurity agency warns there’s an unfortunate imbalance. While attackers can allow AI to act aggressively and accept the consequences when it gets things wrong, defenders can’t. For example, an autonomous security system that mistakenly shuts down critical services could cause more damage than the actual attack. The agency suggests organisations should start by automating low-risk tasks, while keeping humans involved in decisions that could affect live systems.
Work: Moves are afoot to impose controls on the use of personal AI assistants on work devices. Apple device-management specialist Jamf is adding features that will enable employers to detect AI use and, for example, stop staff uploading company files to personal AI accounts. It’s another sign of where workplace AI is heading: organisations won’t simply decide which AI tools staff can use, but increasingly control how they can use them. AppleInsider
In brief
FBI: The agency is investigating a cyber attack that looks like it has exposed highly sensitive information about its staff and investigations. The ShinyHunters extortion group claims to have used a zero-day (previously unknown) vulnerability in Oracle PeopleSoft to obtain names, badge numbers, home addresses and details relating to counter-intelligence and organised crime investigations. The group is now threatening to publish the data unless the FBI corrects a report about it. According to Reuters, the data contains granular detail about sensitive work against Chinese spies, Russian intelligence, and drug cartels. The FBI says the extent of the breach and how attackers got in are still being investigated.
Revolut: Scammers are taking advantage of a recent data breach at the digital bank to try to access more information about affected customers. Malwarebytes says highly convincing phishing texts, some in genuine Revolut threads, are requesting fake identity checks. The breach appears tooriginate from a compromised Italian government email account which enabled the attackers to pose as law enforcement officials requesting confidential user details for ongoing investigations.
Hoist: Hackers managed to breach thousands of companies by compromising the software supply-chain, but were rather less successful at protecting themselves. Google had an undercover analyst inside TeamPCP’s private chats almost from the beginning, while rival hackers later handed over the group’s chat logs. But perhaps the biggest mistake came when one alleged member uploaded stolen material to a Google Drive account linked to his real identity. Google tipped off the FBI; two arrests swiftly followed. WIRED
Dated: Russia has spent years trying to reduce its dependence on Western software. Unfortunate then that an official video from recent elections showed Vladimir Putin voting online using a Dell PC running Windows 10. Perhaps unsurprisingly, an “Activate Windows” watermark suggested the Kremlin hadn’t got round to paying for its copy of an operating system that’s no longer routinely supported. TechRadar
Rude: Sweden has enlisted celebrities to teach people a vital cybersecurity skill: bad manners. Police are urging the public, particularly older people, to overcome their inclination to be polite and hang up on suspicious callers or shut the door on unexpected visitors. Online crime in Sweden raked in 5.7 billion SEK ($579 million) last year, with much of it coming from elderly victims.
Updates
Apple: A new security feature in iOS 27 is designed to detect active social-engineering scams. Impersonation Risk Detection analyses information including download history, purchases and email and phone-call activity, then gives participating apps a risk rating when you’re about to do something sensitive, such as make a payment or change a password. The app doesn’t access the underlying data. You’ll find the feature under Privacy & Security. More mundanely, Apple is working on an iOS 27 update that will fix an issue affecting some iPhone 18 Pro users in which Face ID is failing and forcing a restart.
WordPress: Version 7.1.2 contains a fix for a critical vulnerability that is already being probed by attackers. The fix has also been applied to versions as far back as WordPress 4.7.
D-Link: Two router models are affected by critical vulnerabilities. The older DIR-822A is already out of support and should no longer be used. The newer R95 Wi-Fi 7 router remains supported, but users should check for firmware updates, disable remote administration and ensure the router has a strong, unique administrator password.
Check Point: Update to fix an actively exploited zero-day (previously unknown) vulnerability in Security Management Server.
F5: Update for actively exploited zero-day in BIG-IP Access Policy Manager component.