Cyber Security News in 2026: Theyβre Hacking the Fridges Now!!
![]()
At what point did fridges (refrigerators in 'freedom' speak) become something we needed to defend from a hostile cyberattack?
A quick note before we start: I love AI. I use it, I enjoy it, and Iβm excited about what it helps us build!
What worries me is how easily assistance can become dependence. I find it sad seeing people reach for AI before giving their own ideas a chance. There is something worth protecting in the messy process of thinking, experimenting and coming up with something yourself!
AI can HELP people, and it can be used to HARM people too. This piece focuses on those using it maliciously, and why we need to make a proactive effort to protect ourselves, our businesses and the people who trust us.
Being enthusiastic about AI should include taking its misuse seriously.
Well, here we are... Q4 of 2026. Apparently this is a thing now.
I would LOVE a fridge to remain one of the least demanding relationships in my life. Keep the food cold... make a slightly concerning noise occasionally... I can work with that. I'm a big guy - out of everything I own that hackers could target, this one doesn't sit right with me!
Apparently, it also needs a cyber security strategy.
On 9th November, Reuters reported that around 700 commercial sites across France had been targeted in October's fridge attacks. France's cyber agency reported at least 15 fires. One supermarket owner estimated that the nationwide cost could reach tens of millions of euros.
Fires. In a story about fridges. Literal insanity.
Somebody had to open those shops afterwards and deal with the consequences. The stock was ruined, but the wages and bills hadn't politely disappeared along with it.
The uncomfortable bit is how normal a vulnerable system can look beforehand. A fridge can keep things cold while an unsafe connection sits there unnoticed. Nothing necessarily looks broken. From the owner's side, it is doing exactly what they bought it to do.
Seeing something work tells us very little about what the wrong person could make it do.
That is the thread I keep finding in this year's cyber security news. We see the useful part of a system every day. What happens behind it can receive far less attention.
And AI now lets us reach that convincing, working stage much faster.
How much damage could someone do before you discovered what your own software was capable of?
The problem starts when βit worksβ feels like the job is finished
A founder can describe an idea and get a convincing prototype back surprisingly quickly. The login accepts a password. The dashboard fills with beautiful charts. There is even a button in exactly the shade of green you asked for!
You can end up spending longer arguing with that button colour than checking who can access the information underneath it..
By vibe coding, I mean building through prompts while relying on AI for code you may not fully understand. Tools such as Lovable let founders explore that way of building. I get the excitement! More people can try ideas without waiting for a development budget to appear!
Then the thing starts working, and it becomes very tempting to get it in front of customers...
Consider an imaginary launch conversation:
βDoes the customer portal work?β
βYes! I signed in and uploaded a document.β
βCan another customer read it?β
βI haven't checked that.β
That final answer changes what we have actually demonstrated.
βLet customers upload documentsβ looks like a small request when you type it into a prompt. The business has also acquired decisions about who can read those documents, who can delete them and what happens when access goes wrong.
That is a remarkably small amount of typing for a decision that changes what somebody else is trusting you with.
We can now make a product look ready before we have understood what we are asking people to trust.
Want to see why that question matters? Tech With Tim demonstrates broken account permissions in an AI-built application.
Prefer to open the video separately? Watch the demonstration from 3:24.
A small confession from me...
Before this starts sounding like advice from somebody who always checks everything properly, a small confession.
At Startup Networks, I once found unexpected files on our site. They looked suspicious, so I removed them. Our host later explained that they had been added as temporary protection against excessive automated traffic.
I had removed part of the protection.
Very helpful of me. (I had to go back and apologise profusely)
I suppose the bit I'd be more careful with next time is how quickly βI don't recognise thisβ can turn into βI know what this isβ. Questioning the change was sensible. Establishing who made it, and why, needed another step.
There is a similar temptation when an AI-built product behaves as expected. We start filling in the parts we haven't examined. An integration connects successfully, so we assume it has sensible permissions. The assistant completes a task, so we assume it will stay within that task next time.
Being able to make something happen is a surprisingly persuasive feeling. It can also get ahead of what we understand.
That experience is why I want confidence to come with something we can actually check. Including my own confidence.
Things that do not count as a security audit:
The login works.
Nobody knows the website address yet.
We only have twelve customers.
The AI said everything looked good.
The administrator button is very definitely hidden.
Dave normally deals with this. Dave is on holiday.
Some of these read ridiculous written down. They become surprisingly convincing when there's a launch deadline...
The solution starts before real customer data goes in
There is a point between βwe have a working prototypeβ and βcustomers now depend on thisβ that can be very easy to rush past.
The thing finally works! You want to show people. There is a launch post to write, and stopping to ask awkward questions probably feels like the last thing the project needs.
But somebody is about to trust it with information that matters to them. They cannot see which integrations you connected or what changed during that last round of prompts. They see your business's name and assume somebody has checked.
Their information matters from the first upload however early your business is in its journey.
Keep the experiment small enough to understand. Use sample information and restricted permissions, then get someone competent to review what the application allows against what you actually intended. Understanding the difference between a prototype and an MVP helps frame those decisions. Neither label settles any security questions for you!
Here is the difference between a reassuring demonstration and the evidence I would want next:
The true demonstration | What still needs checking |
|---|---|
A customer can sign in | Each account can access only the information it should |
An administrator button is hidden | The underlying service rejects unauthorised actions |
An integration connects | Its access is limited and its private credentials are protected |
A demonstration usually follows the route you designed. Spend some time checking the routes people should NOT be able to take.
Before pressing PUBLISH, I would want to:
Run realistic tests with separate accounts. Include changing, exporting and deleting information, wherever those actions exist.
Review what each connected tool can do. Establish whether it can read records, change them or send information elsewhere, and whether it needs that much access.
Set approval points for AI. Decide which messages, payments or record changes require a person to check them first.
Check the latest version. Repeat the relevant checks after changes to permissions, integrations or sensitive features.
Name the person who can stop it. Someone needs the authority and access to pause activity, revoke permissions and get technical help.
Use sample data. Finding a missing boundary during testing is a much better conversation than explaining it to the customer whose information was exposed.
Automated scans can help find weaknesses. The business still has to define who should see what, and somebody needs to check that the application follows those rules.
This is also where I would look differently at the time AI has saved. If it takes a week out of the build, does every spare hour really need to go straight back into another feature?
A week saved is also review time made available.
Spend all of it on more features and the checking job is still waiting, with even more software to inspect. A narrower product whose behaviour we can explain gives us somewhere sensible to build from.
The extra dashboard widget can wait. The customer is already here.
![]()
If this is cyber security in 2026, what awaits us in β27!?
My worry for 2027 is how many more existing weaknesses become worth investigating when the cost of trying keeps falling. A business might spend another year postponing the same maintenance, while the people examining it get considerably quicker.
Fresh vulnerability warnings seem to arrive every time we look. Another update. Another exposed service. Another piece of software quietly allowing something its owner never intended.
Finding a weakness gives us an opportunity to fix it. But somebody still needs to understand the warning, establish whether it applies and arrange the work. Meanwhile, the people looking for a way in can keep looking.
Their attempts do not have to stop when we finish work.
And the consequences can reach considerably further than the device somebody first gets into.
On 3rd November, the NCSC warned that AI-powered malware was highly likely to feature in a ransomware attack on Britainβs critical national infrastructure by 2028.
That is close enough for βweβll sort it out next yearβ to stop feeling like much of a plan.
When I think about warfare fifty years ago, I picture tanks, aircraft, ships and soldiers. Food supplies were already targets, with the 1977 Geneva Conventionsβ Additional Protocol I explicitly addressing their protection.
What I struggle to picture is explaining to somebody back then that supermarket refrigeration would become something an intruder could target remotely.
The building could still be standing. The staff could still turn up. The food could still be sitting on the shelves.
And somebody elsewhere could have interfered with the system keeping it safe.
That is what makes the fridge story so much darker than its ridiculous opening suggests. It gives us a very ordinary example of what disruption can look like. You start with an appliance and end up thinking about everything between a delivery arriving and a family eating dinner.
A criminal seeking a ransom and a hostile state seeking disruption both have very different motives. Both can leave ordinary people dealing with interrupted services!
And yes, that scares me.
It also makes preparedness feel much less abstract. On 22th October, the UK government encouraged households to prepare for three days of disruption, including keeping a small supply of bottled water and non-perishable food. It said the approach brought the UK into line with countries including France, Ireland and the Netherlands.
That advice covers emergencies such as severe weather and interruptions to everyday services. Households that can manage temporarily help emergency services concentrate on people who need more support.
I think that is sensible. I also think βdigital resilienceβ sounds rather different once the conversation reaches drinking water and whether the food in your cupboard needs cooking.
Somewhere between discussing software permissions and checking the tins, the consequences become very real.
But personal preparation cannot become an excuse for organisations to keep postponing their own responsibilities.
A cupboard full of tins CANNOT patch the system that failed.
For founders, that responsibility starts when somebody begins depending on what we build. For organisations running essential services, the number of people depending on those decisions can be enormous.
I cannot tell you exactly what 2027 will bring. We have opportunities to improve our defences too. What frightens me is the possibility that we spend another year treating familiar warnings as background noise, while investigating those weaknesses becomes easier.
Because the person responsible for your businessβs security may still be struggling to find a spare afternoon.
The person investigating it can increasingly hand some of their work over.
An attacker can now delegate the boring bits
You have customers to answer, invoices to chase and a task list that appears to reproduce overnight. An attacker investigating your business has work to get through too. AI can help them do more of it.
The NCSC's 2026 assessment identifies AI's use in reconnaissance, social engineering, vulnerability discovery and exploit development. It expects much of the impact to come from accelerating existing methods, with skilled people directing and checking the work.
Anthropic's September report describes observed operations using multiple AI agents for reconnaissance, exploitation and handling stolen information. Humans still set targets and reviewed results. Those are vendor-observed cases, rather than a measurement of all hacking.
What bothers me is the cost of the unsuccessful attempts. Investigating unfamiliar software takes time. Following a promising idea that turns out to be useless takes time. Tools that reduce that work could make more attempts worth pursuing.
βWe're too small for anyone to botherβ becomes a much less comfortable thought when the cost of bothering starts falling.
An open port identifies a reachable service. Whether it can be abused depends on the software, configuration and permissions behind it. AI does not make every attempted exploit successful.
But hoping somebody finds our turnover disappointing feels like a fairly fragile defence. I would rather know why a service is publicly reachable and who is responsible for it.
The attack can arrive looking like an opportunity
A job interview should be a reason to feel hopeful. The possibility that it could become a route into your current employer is a horrible addition to an already stressful process.
Binance's initial account of its October theft, involving about $6.3 billion (YES BILLION WITH A B) in assets, described an engineer targeted through interviews for a fictitious company and a technical assignment containing malicious code. Its wider findings also involved compromised internal AI tools. The investigation was ongoing.
The assignment had a plausible reason to exist. Developers expect technical tasks to involve unfamiliar code.
A business needs protections around what that code can reach on a work device. Everything should not depend on one person recognising a convincing deception while trying to get a job.
The same thinking applies to an urgent payment request or a message apparently from the founder. Verification needs an independently trusted route beyond how believable the message feels. Staff also need permission to use it, especially when the request appears to come from somebody important.
If we tell people to question suspicious requests, then punish them for slowing things down, which instruction do we expect them to remember?
Vibe coding risks can come with a very sleek dashboard
If the green button took three prompts and the permissions took none, I'd have a few questions.
Veracode's autumn 2026 benchmark found that the average security pass rate improved from 56% to 62%. Roughly 38% of tested outputs still contained a vulnerability.
Those numbers would need a bit of care, though. They describe controlled coding tasks, not the percentage of deployed AI-built applications that are vulnerable. The improvement matters. So does the gap between producing working code and producing something ready to handle somebody else's information.
Wiz documented common failures in vibe-coded applications, including authentication handled entirely in the browser, exposed private credentials and inadequate database access controls. These are useful failure patterns to understand, rather than a verdict on every platform or application.
Some protections live underneath the interface. Supabase's row-level security, for example, can enforce rules about which database records an account may access. Those rules still need to match what the business intends.
The useful question at a launch review is who examined that behaviour and what they tested. An attractive screenshot gives us very little help with the answer.
PHP2Shell, and the comfort of βsomeone's dealing with itβ
Meanwhile, the software we already depend on still needs maintaining.
PHP2Shell's two vulnerabilities, CVE-2026-109618 and CVE-2026-109619, can allow remote code execution without a valid login. PHP released fixes on 21st October. An affected system could be made to run an attacker's code remotely.
November also brought Ivanti's warning about a critical Endpoint Manager flaw and a Chrome security update addressing a vulnerability with exploitation reported in the wild. Ivanti said it was not aware of customer exploitation of its disclosed flaws at the time of publication. A critical warning and a confirmed exploited vulnerability tell us different things.
Somebody has to work out whether those warnings apply, arrange the remedy and confirm it happened. There is rarely much glory in that job. βWe checked our PHP versionβ probably won't get the engagement of a product launch post. It still needs doing.
βThe website is up to dateβ might mean the developer updated the application. It does not automatically establish that the host updated PHP underneath it. Two people can give honest answers about different pieces of the same system, while the founder hears that somebody is dealing with it.
Running a business involves relying on other people's work. This is a reassurance worth making specific:
Which version is ACTUALLY running, who checked it, and what changed?
The recovery bill gets paid in evenings
A service can come back online while the people who depend on it are still paying for what happened.
On 13th November, the BBC reported the mental toll on teachers at Bristol schools following a September attack. Thirty-nine schools had lost access to key resources, with some affected teachers saying they were working up to 13 hours a day.
Think about where those extra hours come from.
Lessons still have to be prepared. Children still arrive in the morning. The missing work lands on somebody who already had a full working day, and comes out of an evening they were supposed to have for themselves.
We are far too willing to call that resilience without asking who supplied it.
A small business can experience the same displacement. The founder answers customers individually because the usual system is unavailable. Staff rebuild records. People stay late so customers experience less disruption.
Their effort can make the organisation look healthier than the people keeping it going.
Backups matter. They cannot return spoiled food, undo information already copied or give somebody back the evening spent explaining why an order went missing. Cyber insurance can help address some financial consequences, depending on the policy. Neither removes the work of getting a business functioning again.
That work belongs in the account of what an incident cost. So does the backlog that remains after access is restored. When somebody says βwe're back to normalβ, I want to know whether the people doing the work would agree.
A weekend project can become something another person relies on every morning. Calling it βjust a small appβ will not make the conversation any easier if that person later asks what happened.
![]()
WHY does everything need to be smart?
There is something slightly ridiculous about buying a household appliance and then having to investigate its position on data privacy.
I wanted clean floors. I wasn't expecting homework.
Are we gradually becoming the humans from WALL-E? Handing over another little decision, accepting another convenience, until practically everything around us is doing something on our behalf... and we have very little idea what that involves?
Take ten seconds. Look around the room.
How many things in it have an app, an account or an internet connection? Now pick the one you would struggle hardest to explain.
That is the purchase I want us to think about for a moment.
I can see the appeal of a robot doing the hoovering. Quite a lot of appeal, actually.
But why has βit does more thingsβ become such an easy substitute for βit is a better productβ?
Samsung's own documentation describes features on supported TVs that recognise nearby sounds, including barking and crying, and send notifications through SmartThings.
Still, I would quite like us to pause somewhere around βthe television can recognise the baby cryingβ and ask how many jobs we intended to give the television.
Perhaps that feature is useful to you. Fair enough! But usefulness deserves a conversation about what information it needs, where that information goes and whether you can comfortably say no.
Buying the screen shouldn't feel like agreeing to an expanding job description...
![]()
A fictional product, illustrated for this article. At some point, the extra features became the product, and breakfast became the excuse.
Someone is going to look under the bonnet
People are getting better tools for examining what these products actually do.
REA, short for Reverse Engineer Anything, is a public GitHub project that connects AI agents to tools for inspecting applications and compiled software, including software whose original source code is unavailable. The name is ambitious; it still has technical limits. Helping people investigate software they own is a worthwhile use of AI.
Sometimes curiosity turns up something considerably bigger than expected.
In February, The Verge reported that Sammy Azdoufal used Claude Code while trying to control his DJI Romo vacuum with a PlayStation controller. His investigation exposed access to thousands of other devices. During a demonstration, the reporter watched a colleague's vacuum transmit an accurate floor plan of their home. DJI says it subsequently fixed the underlying permission-validation issue.
He wanted to drive his hoover around for fun.
Other people's homes should never have entered the conversation!
That is the part I find difficult to shake. You can make a perfectly ordinary purchase without appreciating how much access, information and continuing maintenance comes attached to it.
Let 'no thanks' be a feature
For founders, there is a lesson here before the next exciting product meeting. When somebody suggests another connection, another sensor or another AI feature, ask:
What does the customer gain, and what are we asking them to give us in return?
Make somebody explain it. In ordinary language. Including what happens if the customer declines.
And yes, I can hear the objection. Some of these features are useful. For somebody else, the thing I'm mocking might make daily life considerably easier. Automation can give people independence or save work they find exhausting.
That makes the trade-off worth explaining properly. People should be able to understand what access a feature needs, decide whether the benefit is worth it and decline unnecessary access. Useful technology also deserves an OFF switch that doesn't ruin the thing you bought.
I think we overlook the work created by convenience. Someone still has to manage the access, maintain the connection and deal with what happens when it fails. Making a task disappear from the customer's day does not necessarily make that responsibility disappear.
I would like manufacturers to get a little more excited about products that need less access. Less information to protect. Fewer permissions for customers to understand. And occasionally, the confidence to leave a feature out.
Three awkward questions worth borrowing for your next meeting:
What can this access that it doesn't need?
Who checked what happens when somebody uses it differently?
Who can stop it if something goes wrong?
You don't need to turn up with every answer. Asking for one is a useful place to start.
The fridge still has one job
Nobody buys a fridge hoping it will become more involved in their life.
After all this, my requirements remain embarrassingly modest:
Keep the food cold. Let the right people change the settings. Avoid becoming an international incident.
Apparently, that last one needs saying now!
There is an ordinary promise behind most useful technology. The fridge keeps food safe. A business tool lets people get on with their work without quietly exposing them to somebody else's mistakes.
Those promises deserve as much attention as the exciting new features. They are probably why somebody trusted the product in the first place.
We should use tools that help us build. Some of the time they save belongs in understanding what we have built, before someone else has to discover it for us.
Not everything needs to do absolutely EVERYTHING for us. Some things could just do their own job properly.
What is the check you wouldn't skip before putting real customer data into an AI-built app? Tell us in the comments.
Recommended Comments