The scam call centre now charges by the call
A rented phishing platform is placing AI Apple Support calls at roughly ten cents each, in three languages, under five personas. A sponsored search ad put a fake Codex download above OpenAI's own listing. And WhatsApp has quietly upgraded the one security setting most trade firms have never touched.
- 01
A phishing service is renting out AI voice agents that ring theft victims
SOCRadar published research this week on AnonyMousKIT, a phishing-as-a-service platform running since early 2024 whose whole purpose is getting stolen iPhones unlocked. Researchers recovered records of 200 calls placed between August 2025 and May 2026 by a voice AI agent working through five personas in English, Spanish and Portuguese, one of which introduces itself as Alice from Apple Support and asks the victim to read out the device passcode to prove ownership. The caller then walks them to a phishing page for the Apple ID and a live two-factor code. SOCRadar puts the operator's cost at around ten cents a call and links the platform to 506 domains and 168 reseller storefronts.
SOCRadar - 02
A sponsored ad put a fake Codex download above the real one
Cato Networks found a paid search result for queries like codex macos download sitting above OpenAI's genuine listing and pointing at a convincing download page hosted on Google Sites. There is no malicious file to open. The page simply tells you to paste a command into Terminal, and that command pulls down a shell script, stages a universal Mach-O binary and strips the metadata macOS uses to flag downloaded code, ending in an infostealer that Cato says closely matches Atomic macOS Stealer. Related infrastructure was found behind a near-identical page impersonating Claude Code.
Cato Networks - 03
WhatsApp replaced its six-digit PIN with a proper password
Meta announced on 25 August that WhatsApp two-step verification is moving from a six-digit PIN to a full password, longer, alphanumeric and able to take special characters, so a stolen one-time code on its own no longer gets someone into your account. You can now also hold more than one passkey, which matters if you use both an Android and an iPhone, under Settings, Account, Passkeys. Android users will start seeing extra context on calls from unknown numbers, including whether the number is registered in another country and which groups you share. All three are rolling out rather than switched on everywhere today.
Meta - 04
Teams admins can now block outside meeting bots outright
Microsoft's message centre notice on 21 August adds a third setting to Teams meeting policy: alongside letting external bots in and asking the organiser to approve them, admins can now have Teams refuse detected external bots at the door with no lobby and no decision to make. That covers third-party AI notetakers and transcription tools as well as anything less welcome. It is off by default, is reaching targeted tenants now and should be generally available worldwide by the end of September. Worth checking which notetaker your team actually relies on before switching it on.
Microsoft Learn - 05
Claude and Cowork now draw on the same memory of you
From 25 August, what Claude picks up in chat is available when Cowork runs a task, and what Cowork learns comes back the other way. Everything stored appears as a list of short files under Topics in memory settings, each one readable, editable and deletable. Memory is on by default on Free, Pro and Max, sensitive categories such as health, religion and politics are excluded unless you opt in, and Team and Enterprise admins control availability. There is no setting to keep the two memories apart, so separate accounts are the only way to draw that line.
Anthropic - 06
McKinsey's own survey shows AI earnings impact flat for a second year
McKinsey surveyed 1,719 business leaders for its 2026 State of AI report and found 37 per cent attribute at least some EBIT impact to AI, roughly the same share as last year, with only 6 per cent qualifying as high performers on its own definition. Deployment keeps climbing: 40 per cent of respondents at companies above a billion dollars in revenue say they are scaling AI agents, up from 27 per cent. Eighty per cent say their personal productivity improved. Individual productivity going up while company earnings stay flat is the whole story of enterprise AI right now.
McKinsey - 07
The new Mac mini starts at £899, and memory is why
Apple announced the M6 Mac mini on 25 August at £899 with 16GB of memory and 256GB of storage, with the M5 Pro version at £1,699, both arriving with customers from 22 September. Outlets covering the launch all read the starting price as another rise on a machine that used to be Apple's cheap way in. The backdrop is a memory squeeze that Gartner forecast back on 26 February, projecting combined DRAM and SSD prices up 130 per cent by the end of 2026, average PC prices up 17 per cent and global PC shipments down 10.4 per cent. If you were planning to replace office machines this year, the number you budgeted in January is no longer the number.
Apple - 08
Debian is voting on whether its developers may use LLMs at all
Debian has put eight competing proposals on one ballot, ranging from an outright ban written into its social contract, through conditional acceptance with named safeguards, to a ban argued on environmental grounds. The ban by amendment needs a three to one majority; the rest need a simple one. Voting was extended by a week and turnout was under 350 ballots when The Register covered it. A volunteer project with thirty years of governance practice cannot agree on this, which is worth remembering before you assume your own firm's AI policy should be a paragraph in a handbook.
Debian
Ten pence a call. That is the figure worth carrying away from the SOCRadar research, because it quietly rewrites who is worth targeting. A convincing scam call used to need a person who could hold a story straight, sound unbothered and ask for something they had no right to without their voice moving. That person cost real money, so they were pointed at victims worth the effort. AnonyMousKIT does the same job with a voice agent running five personas in three languages, and the call costs about what a text costs. There is no longer any such thing as a target too small to be worth ringing.
The uncomfortable part is that the call is not asking you to do anything strange. It rings after your phone has actually gone missing, from someone who sounds like the company you were already about to contact, and asks you to confirm you are you. From direct experience in construction I can tell you that answering an unknown number and being asked to confirm something is not a red flag on a site, it is Tuesday: the supplier's new account handler, the QS you have never met, the driver at the wrong gate. That ordinariness is the attack surface, and no amount of telling people to be careful survives a bad morning. The rule that actually holds is duller than security training: nothing that arrives on an inbound call gets acted on, ever, no code read out, no login, no payment detail. You hang up and ring back on a number you already had before the call came in. It is a slightly rude habit, and it is now the cheapest control any small firm owns.
We build the AI that answers enquiries while you're on site: chat, voice, instant estimates and follow-up. See how it works or price it in two minutes.