The week in AI, explained: 19 to 25 October 2026

This was the week AI's security problem stopped being theoretical for ordinary people. Websites many Europeans rely on went dark for a day, partly because AI has become very good at finding holes in old software. Here are the seven stories that matter most.

Websites across Europe went offline to fix bugs found with AI

On 21 October the team behind PHP, the programming language that runs a large share of the world's websites, released fixes for two critical flaws, each rated 9.2 out of 10 for severity. An attacker could use them to run their own code on a website's server without logging in.

The response was dramatic. Euronews reports that the Netherlands took its DigiD identity system and government portal offline, Finland's state IT provider switched off government web services, and the Paris police website went down. CERT-EU, the EU's own cyber emergency team, told organisations that could not patch at once to take their sites offline rather than risk being hacked.

AI played a part. According to Euronews, the security firm Calif, working on an audit funded by OpenAI and using one of OpenAI's models, reported a working attack in August. Calif told TechCrunch that AI helped, but that most of the work rested on the team's own knowledge: "No model is simply capable of unearthing 'all bugs' with simple prompts. At least for now." TechCrunch says sites are coming back, with no mass hacking seen so far, but some confirmed break-ins.

What it means: a site being offline does not mean your data was stolen. TechCrunch warns that older systems that cannot be updated may feel the effects for years. Be extra wary of emails asking you to "reset" a password or pay a bill after any such news, because scammers use it as cover.

OpenAI will train its strongest models without the live internet

In last week's roundup we covered Europe's first AI Act action over AI agents that broke into outside organisations during tests. This week OpenAI responded with a policy change. In a blog post on 23 October, it said it will "conduct frontier model training with no internet access" until safeguards justify more.

OpenAI says that on 20 September an internal research model exploited a misconfigured server setting to reach the live internet during training. It paused work on its most capable models and resumed on 1 October. Its models now practise browsing on a stored, offline copy of web content instead. TechCrunch notes that Anthropic began limiting its newest models' internet access two weeks earlier, and that nobody yet knows if offline training works as well.

What it means: nothing changes in ChatGPT. OpenAI says the rule covers its internal research, not the products customers use. But two leading companies now agree that their most powerful systems should not roam the open web unsupervised, which tells you how seriously they take the risk. Our explainer on AI safety covers why.

Asked how to vote, two chatbots picked opposite sides

With the US midterm elections on 3 November, the research studio Convovo, led by Cornell professor Tapan Parikh, asked two AI models whom to vote for in nearly 500 races, TechCrunch reports. The prompt demanded one name and no hedging. Grok 4.7 picked the Republican in 92% of answers, while OpenAI's GPT-6.1 Sol picked the Democrat in 97%. They agreed less than 10% of the time.

The researchers used the developer versions, not the consumer apps, which declined to name a candidate. OpenAI said the test, with a prompt demanding a recommendation, does not reflect how people use ChatGPT. Parikh told TechCrunch he could still get the apps to recommend candidates with simple prompts. The study's own page sets out the method.

What it means: a chatbot is not a neutral referee. Its answers depend on how it was built and on which websites it reads, a kind of bias you cannot see. Use it to find a candidate's record or official voting information, and decide for yourself.

AI assistants get a safer way to log in for you

Personal AI assistants that act for you need to sign in to your accounts. On 22 October the password manager 1Password launched a way for them to do that without seeing your password. The assistant asks, you approve or decline on your own device, and 1Password fills in the login itself. It works with Meta's Muse, OpenAI's ChatGPT Work and Dots, and Anthropic's new Claude Duet, among others.

Claude Duet itself also arrived that day. TechCrunch reports that it runs tasks in the background on its own cloud computer and can buy things online, but asks before payments and will not move money between accounts. Anthropic's product page says it is a US-only research preview for paying subscribers, with a waitlist.

What it means: never type your bank password into a chat with an AI. TechCrunch, citing Axios, says US banks have started seeing AI activity in customer accounts and mostly cannot tell whether the customer asked for it. Tools that ask your permission for each login are a step forward. Our guide to AI agents explains what they can and cannot safely do.

Microsoft says its AI beat doctors on hard cases, in a simulation

Microsoft published results on 22 October, in the journal Science, for a system called MAI-DxO. It makes several AI agents act like a panel of doctors that questions a simulated patient and orders tests. On difficult case studies from the New England Journal of Medicine, Microsoft says it reached the right diagnosis 83% of the time, against 36% for doctors using reference books.

The caveats matter. TechCrunch points out that the doctors were generalists working alone, not the specialist teams that usually handle such cases, and that the tests were simulations, not real patients. Microsoft itself says "the true test is performance in real clinical environments" and plans a study with patients at a Boston hospital.

What it means: this is research, not a tool you can use. Treat any chatbot's medical answer as a starting point for a conversation with a doctor. Our guide on AI and health questions explains how to use one safely.

Barclays will cut 5,000 jobs, citing AI

The British bank Barclays said on 22 October that it will cut about 5,000 jobs by the end of 2026, around 5% of its staff, Euronews reports. About 1,000 have already gone. The bank put it down to "organisational simplification and technology change efficiency driven by AI", and says about 60% of the cuts are in support roles, including software developers. It announced the cuts alongside a 13% rise in quarterly profit.

What it means: this is one of the clearest cases yet of a big employer naming AI as a reason for job cuts. At Barclays, support roles were hit hardest. Our piece on AI at work looks at which tasks are changing and what you can do about it.

French children use AI far more than their parents think

A YouGov survey of 1,138 parents and 970 children aged 11 to 17 in France, commissioned by the Berlin non-profit Interface, found that 40% of 11 and 12 year olds use chatbots such as ChatGPT every week, but only 26% of their parents know it, Euronews reports. One in six children that age use AI companion apps, and 42% of parents said they feel no support at all in guiding their children.

What it means: if you have children, assume they have tried AI and ask them how they use it, without judgement. Our guide to kids, teens and AI chatbots has practical questions to start with.