The impact of AI discovery on vital infrastructure. Cybercrime gets real.
Bosses break security basics. Lies, damned lies, and marketing folk. And Meta's drive for more revenue.
All change
Last week, we wrote about the looming crisis posed by the growing volume of security vulnerabilities and the limited ability of organisations to address them. This week brings more evidence pointing to the scale of the problem. Six weeks after launching its Mythos cybersecurity model, Anthropic says researchers and partners using it have found more than 23,000 vulnerabilities across more than 1,000 open-source projects. Many security researchers have dismissed Anthropic’s announcements as heavy on marketing hype and short on detail. While it’s obviously in Anthropic’s interest to talk up its capabilities, the sheer number of vulnerabilities (many them serious) does point to an existential crisis for the existing technology ecosystem.
Many of the 1,000 open-source projects analysed by Mythos underpin modern technical infrastructure, including the internet (and Anthropic’s own technology). It’s well-documented that such projects are often underfunded and dependent on the efforts of volunteers. Such a system is completely unprepared and incapable of dealing with the volume of vulnerabilities being identified by AI models like Mythos. As Anthropic points out, similar models to its own are soon going to be available more broadly. Anthropic’s solution, put simply, is for software developers to work harder and network defenders to work faster. While AI can help mitigate the challenge, the reality is that the technology on which modern society is based was not designed with AI capabilities in mind. The effects of the mismatch will not be pretty.
Threats
Personal: The FBI says a ransom group is now sending people to offices if phishing attempts fail. The gang impersonates internal IT staff and tries to persuade victims to grant remote access. If that fails, the FBI says someone may visit the target in order to connect a USB stick or external hard drive. In a similar vein, Bloomberg reports that the cryptocurrency industry is stepping up security measures in the face of kidnappings, assaults and armed home invasions. A recent conference in Las Vegas saw speakers protected by bodyguards and workshops focused on what holders of cryptocurrency should do in the event of a home invasion.
Vault: We don’t recommend Google’s Password Manager, but if you use it you should be sceptical of requests to enter your six-digit PIN - even if they appear to come from Google. Researchers have identified a technique that subverts the authentication process and presents users with an all too realistic copy of the real prompt. Double-check before entering your PIN; this attack means that, if stolen, it gives access to everything in your vault. PhishU
Threats: Victims of phone theft in London are increasingly receiving threatening messages after their devices are stolen, with criminals trying to intimidate people into revealing passwords or removing security locks. The messages often claim the phone has been found or warn that personal information will be leaked unless the victim cooperates. New York Times
Texts: A widespread campaign is impersonating government portals, telecoms, and parcel delivery services across 19 countries. It uses text messages to trick victims into handing over credit card information that can be sold on underground hacking portals. The campaign’s main focus has been the UK and Ireland. Hunt
Balls: It’s no surprise that scumbags are seeking to exploit the 2026 FIFA World Cup, but the scale of their efforts is impressive. Researchers have identified thousands of suspicious domains linked to ticket scams, fake streaming services, counterfeit merchandise and phishing attacks impersonating sponsors and travel providers. Criminals are also using AI-generated content and convincing fake websites to target fans seeking tickets or accommodation. Flare
Surveys: KnowBe4 has details of a large-scale scam campaign using fake customer surveys and free gift offers to harvest credit card details and personal information. The attacks impersonate well-known brands including Costco, FedEx and Marriott, using convincing websites, countdown timers and fake customer comments to pressure victims into handing over payment details for supposed “shipping fees” or prize claims.
AI
C-suite: Based on our experience, it’s no surprise to read that senior executives are among the worst offenders when it comes to using unauthorised AI tools at work. In a survey by TrustedTech, nearly two-thirds of decision-makers admitted using “shadow AI” despite understanding the associated security and privacy risks. That’s double the rate of ordinary employees. The findings highlight a growing disconnect between AI policies and real-world behaviour, particularly as staff increasingly turn to unsanctioned tools to save time or evade restrictions.
Search: Internet search alternative DuckDuckGo says it’s seen a 30% increase in installs in the US since Google revamped its model to make AI unavoidable.
Creaky: Researchers have found a serious vulnerability in a widely used but little-known component found in many AI server products. The flaw could allow attackers to bypass authentication controls and gain access to private internal services by tricking servers into treating malicious requests as trusted ones. It’s a reminder that amid the pell-mell rush to adopt AI, security is by no means guaranteed. X41 D-Sec
Proactive: Prepare for hell. Customer service company Talkdesk has unveiled AI call centre agents designed not just to respond to customer queries, but to proactively contact people before they ask for help. The systems are aimed at retailers and banks, and will automatically follow up abandoned shopping carts, push loan applications and chase debts. Call screening is about to become essential.
Trump
Order: President Trump almost signed an executive order on artificial intelligence oversight last week. Almost, but not quite. Politico has obtained a copy of the order which was designed to address concerns about the possible impact of AI misuse. Despite the contents being voluntary, the order remains unsigned with the president expressing concern it might slow US competition with China.
Phone: Yes, that phone. While Trump Mobile managed to send devices to some reviewers, it doesn’t seem to have dispatched any to actual customers. What it has managed to do is expose their details (all 27,224 of them) thanks to slipshod security on its website. Guardian
In brief
Liars: Long-time readers might recall a story from 2023 about a US media company that claimed it could use smartphones to listen to people’s conversations in order to target adverts. We were sceptical about their claims, as it turns out, with good reason. This week, the US Federal Trade Commission fined Cox Media and two marketing partners $930,000 to settle allegations that the claims were false. Cox told potential clients they could ensure “every casual conversation between two consumers becomes a tool for you to target, retarget, and retain customers.” The FTC found the supposed service was completely fabricated. As we’ve always said, it often feels like technology companies are listening to us. That’s because they are, just not to what we’re saying. They get all the information they need from all the other ways they monitor our online activity.
Tracked 1: US military officials have confirmed that hostile states are using commercially available location data to target troops in operational areas. The confirmation came in a letter from US Central Command whose area of responsibility includes the Gulf. Location data is fundamental to digital advertising and is collected from smartphones and other devices before being flogged, sometimes via complex networks of intermediaries. Reuters
Tracked 2: If you want to be anonymous online, it would be a good idea not to use your own photo in profiles. The FBI was able to identify a prolific purveyor of AI-generated sexual deepfakes by connecting his profile photo to social media activity. In another case, they identified a suspect with some pretty simple cross-referencing of online activity and an IP address. DoJ
Tracked 3: You might have seen the news that a German court has sentenced a former member of the Baader-Meinhof gang to 13 years in jail. She had evaded police for more than 30 years, so it’s worth recalling how she was found because it underlines the power of facial recognition software (which German police aren’t allowed to use). It took a Bellingcat researcher about 30 minutes to track her down using old photographs and a couple of publicly available services. These linked the old images to current ones of the wanted woman in Berlin.
Storage: A new browser-tracking technique can identify what websites and apps people are using by analysing what their computer’s solid state storage is doing. It works by analysing subtle timing variations in disk activity and appears capable of bypassing anti-fingerprinting protections. While technically impressive, its real-world usefulness is probably limited outside highly controlled scenarios. Ars Technica
Subscriptions: Projections put Meta’s global advertising revenue for this year at $243.5 billion, but that’s clearly not enough because the company has just started rolling out paid subscriptions for Instagram, Facebook and WhatsApp. Meta says the new plans are “just the beginning with a lot more value to come.” TechCrunch
Updates
WhatsApp: A “zero-click” attack targeting older iPhones running iOS 16 allowed attackers to hijack WhatsApp accounts without any interaction by the victim. The campaign appears to exploit vulnerabilities patched in newer iOS releases, underlining the need to apply updates as soon as possible. Forenser
Microsoft: In the face of, shall we say, ‘negative’ customer feedback, Microsoft is giving users the option of making the Copilot button a little less intrusive. Forums have been filled with “infuriated” complaints that the floating button was getting in the way of actual content. There’s now a simple way of banishing the button to the toolbar.
HP: An auto-installed update has left some of the company’s most expensive computers unusable. The problem affects the devices’ BIOS software which underpins their complete operation. HP has provided a multi-step workaround likely to drive users and administrators slightly crazy.
Ubiquiti: Security updates on Friday to address five vulnerabilities in its UniFi operating system. Three of the five issues have a severity rating of 10/10 and can be exploited to takeover devices.
Zimbra: Version 10.1.17 includes important security updates as well as improved functionality.