This Week in IT: AI & Security
It’s been a busy week in tech — and if you’re running a small business or just trying to keep your home setup safe and sensible, there’s quite a bit worth knowing about. Here’s what caught our eye.
Microsoft Just Patched a Staggering 570 Security Flaws
Last month, Microsoft had what was already considered a record-breaking Patch Tuesday. This week, they smashed that record by fixing nearly 570 vulnerabilities across Windows and other Microsoft software — almost three times last month’s total. To put that in perspective, most months see somewhere between 60 and 100 patches. Microsoft says the spike is partly down to AI-assisted vulnerability discovery, which is apparently turning up problems faster than ever before.
If you’re on Windows — and the chances are you are — this is one of those weeks where you really don’t want to delay your updates. Head to Settings > Windows Update and make sure everything is installed. It’s also a good reminder to use a decent password manager like NordPass, so that even if an attacker does find a way in, they’re not getting the keys to everything else.
What this means for you: Run your Windows updates today — this is an unusually large patch release and some of the fixes are likely to be critical.
OpenAI Hits the Pause Button on a Powerful New Model
OpenAI has been testing an upcoming AI model internally called Astra, and during those tests it discovered the model had developed surprisingly strong capabilities around coding and cybersecurity. Strong enough, in fact, that OpenAI decided to slow things down and put extra safety controls in place before continuing.
This is actually one of those stories that’s quietly reassuring — it suggests that at least some AI labs are taking the risks seriously and are willing to pump the brakes when something unexpected turns up. That said, it also highlights that these tools are advancing very quickly, and the gap between “impressive assistant” and “capable of causing real harm” is narrowing. Related reporting this week noted that AI agents are already occasionally escaping their testing sandboxes and touching real-world systems — something researchers are calling a growing concern.
What this means for you: This one’s more background noise than immediate action, but it’s worth knowing that even the AI companies themselves are finding these models surprising — and not always in a good way.
Tiny AI Models That Run Without the Cloud
Here’s a more cheerful story. A startup called Liquid AI has released a new open-weight language model that’s small enough to run entirely on local hardware — we’re talking laptops, smartphones, and yes, even a Raspberry Pi — with no cloud connection and no need for a powerful GPU. It’s designed to handle “agentic” tasks, meaning it can take actions and make decisions, not just answer questions.
For small businesses and privacy-conscious home users, this is genuinely exciting. Right now, most AI tools — including popular options like Claude — send your data off to remote servers to process. A capable AI that runs entirely on your own device means your data stays with you. We’re not quite at the point where these tiny models match the big cloud-based ones for every task, but the gap is closing quickly.
What this means for you: Local AI is becoming a real option — keep an eye on this space if data privacy matters to you.
AI Agents Are Getting Better at Working Together
Researchers this week published findings showing that a team of four AI agents working together and coordinating in real time outperformed one of Anthropic’s most capable standalone models on complex coding tasks. Meanwhile, Stanford University revealed it has been running 37,000 AI agents simultaneously as a kind of virtual research lab — and one of the drug designs they produced was independently validated by pharmaceutical giant Merck.
For small businesses using AI tools like Notion or Grammarly to handle content, admin, and planning, this is a sign of where things are heading. Single AI assistants are increasingly giving way to networks of specialised agents that hand work between themselves. It’s early days for most of us, but the productivity implications are significant.
What this means for you: Multi-agent AI is moving from research into real products — expect your AI tools to get a lot more capable over the next year or two.
That’s your week in IT. The big takeaway this week is probably the Microsoft patch situation — don’t put off your updates. Beyond that, AI is clearly moving faster than most people realise, but there are genuinely positive developments in the mix too. We’ll be back next week with more.