Today's briefing
The scam call got cheap before the software got good
Interpol ties over half of one region's cybercrime to AI, Microsoft admits the Copilot nagging is deliberate, and your blog's AI header image is costing you rea
The one that matters
Interpol: AI now fuels most cybercrime in one region. The same scam kit works anywhere
What happened, in plain words
Interpol reports that AI is now behind more than half of cybercrime in Africa, with scams in particular surging. The finding is regional. The tools are not. The same voice cloning, the same fluent email in your local business idiom, the same convincing fake invoice that works in Lagos works just as well in Ludhiana or London, Ontario. What changed is not that criminals got smarter. The expensive part of fraud — sounding like a real person who knows your business — has become close to free.
What it actually means for your business
For twenty years, staff were trained to spot fraud by how it looked: broken English, odd formatting, a sender address that was obviously wrong. That training is now worthless, and worse than worthless, because it gives people false confidence. A well-written email is no longer evidence of anything.
Think about where this lands in your own operation. A 600-student school whose fee clerk gets a WhatsApp message from the principal's number-that-looks-right, asking to redirect a vendor payment before the audit. A 30-bed clinic that receives a polite letter from its regular oxygen supplier updating its bank account details. A workshop with 40 staff where the owner is travelling and a voice note arrives — the owner's actual voice, ten seconds long, slightly rushed — telling the accountant to release a payment today.
Every one of those attacks targets the same thing: your payment approval process, not your computers. And the loss is not abstract. One wrong supplier transfer at a 40-person workshop is typically a month of payroll. Most small firms never recover it.
What it does NOT mean
It does not mean you need to buy an AI security product. Expect a wave of vendors quoting this exact Interpol finding in sales decks over the next few months — that is the predictable move, and it is mostly noise. The report describes the volume and cheapness of AI-assisted crime, not attacks that defeat basic verification. Nearly all of this is ordinary fraud with better spelling and better timing.
It also does not mean your business is more likely to be singled out. You are not being targeted personally; you are being targeted cheaply, at scale, alongside ten thousand others. That is actually good news, because generic attacks are stopped by generic discipline.
What to do this month
- Write one rule and post it at the accounts desk: no bank account details are ever changed on the strength of an email, a WhatsApp message or a phone call. Only after a callback to the number you already had on file — never the number in the message.
- Require two people to approve any payment above a number that would hurt. A week's payroll is a sensible line for most firms.
- Tell staff plainly that a familiar voice is no longer proof of identity. Say it out loud, in a meeting. The person most likely to be hit is your most junior finance person, late on a Friday.
- Agree an internal phrase for urgent payment requests that only your senior team knows.
Total cost: an hour of your time and a printed sheet. There is nothing here to buy, and any consultant who tells you otherwise is selling to the headline rather than the problem.
Also worth knowing
-
Six outlets are now asking out loud whether AI demand is real
The question being raised is whether the enormous spending on AI capacity reflects real customer demand or circular deals between a handful of large firms. For you this is not a stock tip, it is a contract question. Assume today's pricing is promotional, keep your data exportable, and do not sign a three-year lock-in for a tool you have used for three weeks.
Hacker News ↗ -
Microsoft confirms the endless Copilot prompts in Microsoft 365 are deliberate
If your staff keep bumping into Copilot upsells inside software you already pay for, that is the design working as intended, not a bug that will pass. Decide deliberately whether you want Copilot seats at all, then use tenant admin settings to switch the prompts off for everyone who does not — rather than telling people to ignore them.
Windows Latest ↗ -
A live supply-chain attack has hit widely used npm packages, including Keyv
Malicious code arrived through ordinary software updates rather than anyone breaking in — the developer equivalent of a poisoned delivery from a trusted supplier. You almost certainly do not touch npm, but your web developer or agency does. Send them one email this week asking whether anything they maintain for you pulled a compromised version, and when your site's credentials were last rotated.
Hacker News ↗ -
AI-generated header images are making readers close the tab
An argument that AI-generated blog images actively put people off reading drew 727 points on Hacker News, which tells you where taste has landed. The cheap synthetic banner now signals "nobody spent any effort here," which is the opposite of what you want on a services page. A real photograph of your own workshop, clinic or classroom costs nothing and cannot be mistaken for filler.
Hacker News ↗ -
OpenAI ships education tools for teachers and students
New education plugins for ChatGPT Work and Codex are aimed at K–12 teachers, college educators and students. If you run a school or coaching centre, the near-term value is teacher time — lesson plans, question banks, first-pass marking — not student-facing tutoring, which still needs supervision. Pilot with two willing teachers for one term and count hours saved before buying seats for the whole staff room.
OpenAI ↗ -
Another strong coding model lands, and the price of building software keeps falling
Qwen3.8-Max topped Hacker News with over a thousand points, continuing a steady pattern: capable coding models keep getting cheaper. This matters to you indirectly but concretely — it is why a quote for a small internal tool, a booking screen or a stock tracker should be falling year on year. If your vendor's number looks like 2023 pricing, it is fair to ask what changed.
Hacker News ↗ -
A small model built to run on your own device, not someone else's cloud
LFM2.5-2.6B is designed to run locally — on a laptop, a phone or a shop-floor machine — rather than sending data out. For a clinic, an accountant or anyone handling records under India's DPDP rules or Canadian privacy law, that direction matters more than any leaderboard position. It still needs a technical person to set up, so treat it as a question for your IT vendor in six months, not a purchase this week.
Hugging Face ↗ -
A crypto site asked Copilot to predict a token price. This is not a thing chatbots do
A chatbot does not forecast markets; it produces text shaped like a forecast, with the same confident tone whether it is right or nonsense. Anyone quoting an AI price prediction — a vendor, an adviser, an enthusiastic nephew — is telling you something about their judgement, not about the market. File it next to horoscopes and move on.
Cryptonews ↗
How this briefing is put together
Every morning we read the day's AI announcements and reporting from the companies themselves and from the technology press, then pick the handful that actually change something for a working business. The analysis is ours and it is written for owners and managers, not engineers. Every story links to its original source above — read them, and disagree with us where we've got it wrong.