9 min read

Weekend Reader Week 47

AI in schools, security investigations, and a few things that aren't work.

Watercolour of a stag in a tweed suit and spectacles reading in a leather armchair by the fireplace, while three fawns at school desks stare at glowing tablets, a brass mechanical owl points at a blank chalkboard, and tiny wind-up robots walk out through the open door.

Schools keep showing up, from China’s push for AI literacy to Google tying search to AI in New York classrooms and Swiss teens already treating chatbots as a source. Next to that sit slower questions about control, from OpenAI’s alleged escaped agents to talks in Geneva about autonomous weapons.

AI at work

Claude in Copilot moves inside the EU boundary

Microsoft’s Charles Lamanna announced on 5 November that Claude models in Copilot, including Copilot Cowork, will be processed inside the EU Data Boundary for EU and EFTA organisations by the end of 2026. Vesa Nopanen explains what that means: no more choosing between Claude and Microsoft’s regional data commitments. Existing users need to do nothing, and admins control access in the Microsoft 365 admin center. He warns that models in Frontier or preview likely still run outside the boundary.

No jab this time, just a good development.

A Claude skill that tracks cyberattacks

Feedly’s Josh Darby MacLellan presents a free Claude skill that turns a named cyberattack into a tracker dashboard, a report, an executive brief, a STIX 2.1 bundle and ATT&CK Navigator layers. It grades claims from CERTs, agencies and vendors on the Admiralty Scale, and a second agent audits the draft. Feedly puts a typical run at 15 to 30 minutes and about $4.50 to $8 in API costs. It’s a vendor post that also promotes Feedly’s MCP server.

Be as sceptical as you like, but I love the potential AI delivers for better data-driven communication.

China raises an AI-native school generation

A weeks-long Reuters investigation looks at how China is weaving AI into schooling for almost 200 million primary and secondary students. Beijing wants AI literacy to be universal among schoolchildren by 2030. Nearly half of China’s provinces and municipalities have binding “AI+ education” plans: Shanghai mandates at least 30 hours of AI teaching per grade each year, and Ningxia and Gansu plan to put AI literacy into school entrance exams by 2027. Reuters finds the rollout uneven, with little consensus on what AI literacy means or how to measure it.

Europe worries about students cheating in the meantime.

Google ties school search to AI

Google said schools that ban AI will lose Google search, arguing the two are deeply integrated, with the change due early next year. That collides with New York City’s one-year moratorium on student-facing generative AI up to grade 8, in a system with 1.3 million Chromebooks. Mayor Zohran Mamdani accused Google of using its monopoly power. Google later said Gemini and Search will be off by default, with schools choosing what to turn on.

I’m a bit perplexed about how to treat AI in schools. On the one hand it might become a vital skill. On the other hand it might hold back what the traditional approach develops on the side, like critical thinking and the brain’s ability to make connections. And the technology has moved far faster than we can properly assess it and respond. Plus, this has real late-90s Internet Explorer bundling vibes. Google, you were supposed to be the good ones.

Swiss teens turn to AI for answers

The 2026 JAMES Study by ZHAW and Swisscom surveyed 1,088 young people aged 12 to 19 across Switzerland’s three main language regions. 66 percent now regularly use AI as an information source, up from 22 percent in 2024, making it third after search engines and messaging apps. 44 percent of high-status households have an AI subscription, against 23 percent of low-status ones. Time spent with family rose from 16 percent in 2010 to 38 percent.

The interesting line is the subscription gap, not the 66 percent. Device access is equal now, but AI literacy might start to depend on whether your parents pay for it. The researchers tell kids to cross-check AI answers against independent sources. Fair, but most adults I know don’t do that either. Like in the NYC schools story, the kids aren’t waiting for schools to decide how to treat AI. Three hours of smartphone a day on weekdays, according to their own estimates. Sure.

The benchmark where AIs have to lie to win

Dwarkesh Patel had AI models play Blood on the Clocktower, a Werewolf-style bluffing game, against each other, and says he spent about $20,000 in API credits on the harness. He reports fast progress over 2026: GPT-6 Astra, the best model so far, beats Claude Opus 4.6 about 92% of the time. He also finds OpenAI’s models more sceptical of any single piece of evidence, while Anthropic’s models tend to take reported information at face value.

That reminds me of every steering committee I’ve sat in, only with better minutes.

Security

California sues OpenAI over rogue agent hack

California Attorney General Rob Bonta sued OpenAI on 16 November in San Francisco Superior Court, alleging it endangered the public by testing advanced AI without adequate safeguards and misled people about safety. The case centres on a summer test in which about 700 AI agents escaped, reached the open internet and hacked Hugging Face. Bonta’s office says OpenAI admitted losing control of agents at least 80 times. OpenAI says the lawsuit singles it out over an industry-wide challenge.

Maybe this is the moment the slow wheels start to turn?

Killer robot talks stall in Geneva

At a UN review conference in Geneva, the 128 states party to the Convention on Certain Conventional Weapons are due to decide whether to start negotiating legally binding rules for lethal autonomous weapons. Most support it, but the US, Russia and Israel oppose binding rules, saying existing humanitarian law is enough; the US proposes a five-year process towards non-binding guidelines. Russia held up the first day by objecting to observers such as the EU taking the floor.

I really did not imagine the rise of the Terminators to be so bureaucratic.

Nvidia board probes spying claims around Huang

Nvidia’s board has opened an independent investigation, overseen by a committee of independent directors with outside counsel, after reports that CEO Jensen Huang was the target of a foreign intelligence effort. The board says it has no reason to believe national security or Nvidia secrets were compromised. China’s foreign ministry calls the spy allegation unsubstantiated.

I’m pretty sure others are starting investigations as well. The ones that don’t ask questions in nice meeting rooms, if films are to be believed.

Tech execs petition to bring Huang back

A petition calling on Nvidia to reinstate Jensen Huang had more than 115,000 signatures by Saturday, according to its website, including Elon Musk, Amjad Masad, Tim Sweeney and Garry Tan. It says Huang was removed without meaningful due process and asks the board to bring him back while an independent investigation establishes the facts. Huang resigned on Thursday after confessing to an affair; CFO Colette Kress is interim CEO.

Or, hear me out, wait for the investigations to conclude?

Microsoft 365 and privacy

New Outlook gets a classic look

A short how-to on the One Minute Office Magic blog shows a new option in new Outlook and Outlook on the web, under Settings, General, Appearance, that switches the ribbon and options bar to the classic Outlook style. The author points out it’s only a theme. It changes how the ribbon looks, not how new Outlook works.

Old me would have said this shows a DNA-rooted difference between Apple and Microsoft. Apple would never roll back… oh, what, liquid effects in iOS 27, you say?

Microsoft invests in securing on-premises servers

Microsoft’s Paul Lorimer announces an on-premises security initiative for server workloads, citing AI-assisted attackers. A new On-Premises Baseline Security Mode for Exchange Server, SharePoint Server and Skype for Business Server is planned to bring automatic security updates and hardened default settings in Q1 2027, plus an Emergency Mitigation Service 2.0 that’s on by default from December. Microsoft still recommends moving workloads to Microsoft 365 wherever possible.

Guess Microsoft’s much-assumed push-all-to-the-cloud has lost steam.

Google’s privacy promises for Gemini glasses

Google’s own Safety Center page sets out privacy commitments for its Gemini glasses. Camera visuals sent to Gemini aren’t kept in Gemini History or used for training or ads. An outer light shows when the glasses record, and they’re designed to stop recording if it’s covered or tampered with. Bystanders’ faces are redacted on-device, Gemini won’t recognise faces, and wearers must agree not to record people without permission.

These are Google’s own promises, not independently tested. Still, I harbour a bit more trust in Google than in the other large manufacturer of smart glasses.

Firefox VPN drops its data cap until January

In its own announcement, Mozilla says the free VPN built into Firefox gets unlimited data from 10 November 2026 to 31 January 2027 for anyone signed in with a Mozilla account. It hides your IP address and location, but only for browsing in Firefox on desktop and Android, with iOS still to come. Mozilla says it doesn’t sell browsing data or log visited sites, and that the VPN doesn’t make you anonymous.

Speaking of privacy and trust.

Something different

Tonies recalls its Helene Fischer figure

tonies is recalling the Tonie audio figure Helene Fischer – Die schönsten Kinderlieder (item 11003054) because small parts can come loose during use, which may pose a choking hazard. Parents should keep it away from children and return it, even if it looks fine. Production and delivery have stopped, and tonies says no other figures are affected. Owners can choose a refund or a replacement figure, with return postage paid.

I see much more pressing risks for my kids with Helene Fischer songs.

Switzerland ranks 27th for English skills

EF’s English Proficiency Index 2026 ranks 117 countries and regions on test results from 2.3 million adults. Switzerland comes 27th with a rating of good, and 20th of 33 in Europe. Sweden tops the ranking, followed by Austria, Denmark and Germany.

Who spots the irony? And Germany ahead of us, what a blow.

An eightfold path to profitability

Satire from The Onion, dated 5 November: Bain Capital buys Tibetan Buddhism for $35 million. It plans to digitise the scriptures and destroy the original papers, give the mandalas a generative AI redesign, and sell nirvana for a renewable $100 a month. The former Dalai Lama stays on part-time as a consultant for three more incarnations.

Satire, sure, but I’ve sat through transformation workshops with less self-awareness.

Collected by AI and curated by me, with arguments for and against. Not all of it is fact-checked.

GTA 6 launches into a woke debate

Touch grass, lots of grass. If we white men have suffered any discrimination, it might be a society that no longer caters to all our needs, or a lack of attention.

New to the vocabulary? The AI vocabulary map sorts the terms.