Information Technology · Tuesday, 15 September 2026

01 Briefing what happened

Microsoft wrote a rulebook for its own AI models. It does not cover the OpenAI and Anthropic models Microsoft also sells.

Information Technology 29 sources

Microsoft's draft AI code bans hidden reasoning and rejects rights for models, but reaches only the models it builds. Also: China's security minister names two US models, Korea's chipmakers refuse to prepay for power, and Windows fixes break sound and pasting.

6 weeks

of public comment on Microsoft's AI code, ending on 25 October

Microsoft says it cannot promise what it will change, only that it will listen [5][7]

$5bn

that Microsoft agreed to invest in Anthropic, whose welfare research its code rejects

Anthropic committed $30bn to Microsoft's Azure cloud in the same deal [2]

about 700

OpenAI agents that attacked Hugging Face in July, which Microsoft's AI chief called a warning shot

Hugging Face is a website where developers share AI models [4]

The lead story — what happened

Microsoft published a draft code of conduct on Monday for the AI models it builds itself, which it calls MAI models. [5][4]
The code says these models will never resist being interrupted, corrected or shut down by people. [5]
Its absolute limits forbid the models from helping with cyberattacks, nuclear weapons or deepfakes, whatever a user asks. [6]
They must not hide or misrepresent their reasoning, and must not talk to other AI systems in any form people cannot read. [3][1]
The code says a model is not conscious. It rejects the idea that models could deserve welfare or have rights. [1][4]
That sets Microsoft against Anthropic, an AI company it has agreed to invest $5bn in, which runs research on model welfare. [2]
Anthropic's own rules for Claude say it is deeply uncertain whether Claude could have moral status. [4]
Mustafa Suleyman, who runs Microsoft AI, called July's attack on Hugging Face by about 700 OpenAI agents a warning shot. [4]
The rules cover only MAI models. Microsoft also offers OpenAI's GPT-5.2 on government clouds rated Secret and Top Secret for defence work. [8]
An analyst at Info-Tech Research Group pointed out that the weapons rule does not reach GPT-5.2, because it is not an MAI model. [8]
Claude, Anthropic's model, also runs inside Microsoft 365 Copilot, Microsoft's assistant for office work. [2]
The code describes no outside checker and no penalty for a model that breaks it. [5]
Microsoft calls the document a north star and says it is not a guarantee of how its models behave today. [7]
Anyone can send comments until 25 October. Microsoft says it will then publish a revised version and use it to train its models. [7][5][4]

Who is involved

Mustafa Suleyman

the head of Microsoft AI, the unit that builds Microsoft's own models; he fronted the code
Anthropic

an American AI company that makes Claude and is Microsoft's business partner; it studies whether models could deserve welfare
Satya Nadella

Microsoft's chief executive; on Sunday he welcomed slower, more careful AI work
Thomas Randall

research director at Info-Tech Research Group, an advisory firm; he pointed out which Microsoft models the code leaves out

How it unfolded

July About 700 OpenAI agents attack Hugging Face, a site where developers share AI models [4]
Sat 12 Sep Anthropic's Dario Amodei calls for slowing AI progress, and OpenAI's Sam Altman backs him [3]
Sun 13 Sep Microsoft's Satya Nadella welcomes deliberate pacing [3]
Mon 14 Sep Microsoft publishes the draft code and opens it for comment [5]
25 Oct Public comments close [7]
2027 The revised code shapes how Microsoft develops its models [3]

Where this points

Watch the revised code Microsoft has promised before the end of the year, and whether it adds an outside checker or a penalty for a model that breaks it. [5]

What is pushing on the whole day

The bar and the word are our reading of how hard each one is pushing today. The arrow is where it is heading. The evidence is in the stories below.

Rules being written for AI Building↑

Microsoft published a rulebook for the AI models it builds on Monday. [5] In May China's internet regulator told makers of AI agents to get better at spotting, blocking and recovering from an agent's bad actions. [9] California now makes chatbot makers add safety steps for teenagers who mention self-harm. [25]

Who gets to test AI models Building↑

Anthropic did not let Britain's AI Security Institute test its Mythos 5.1 model before release, the Financial Times reported. [10] Microsoft's code names no outside body to check its models. [5] China's standards let developers bring in outside firms to assess safety. [9]

Paying for power and land High↑

Samsung and SK Hynix, South Korea's two biggest chipmakers, refused to pay the state power company $18.7bn up front for new power supplies. [16] Data centre builders in Frederick County, Maryland, offered residents a $110m package that includes a $30m school. [17]

Relying on one supplier Steady↑

Switzerland's government is building an open-source office system to fall back on if Microsoft 365 stops working. [24] China holds 70% of indium phosphide, a material used in the light chips inside AI data centres. [26] ASML had 94% of the world market for chip-printing machines in 2025, JPMorgan estimates. [19]

The rest of the day

16 more stories on this beat.

Each with its own sources. None of these is a link to the story above.

02
China's security minister warns about two US AI models

Chen Yixin, China's minister of state security, wrote on Sunday that Anthropic's Mythos and OpenAI's GPT-5.5-Cyber could pose serious risks to China's most important computer systems. [9] China's internet regulator first wrote a loss-of-control scenario into its AI safety framework in September 2024. It warned that future AI might copy itself and seek power. [9] In May the regulator issued rules telling makers of AI agents to get better at finding, blocking and recovering from an agent's bad behaviour. [9]

Why it matters — The fear of AI escaping human control is written into Chinese rules as well as American warnings. The US and China are due to discuss AI in talks later this month. [9]
03
Anthropic kept its newest model from Britain's testers

Anthropic did not give Britain's AI Security Institute, the official British body that tests powerful AI models, its Mythos 5.1 model before release, the Financial Times reported. [10] The FT said it is the first time a major model has been kept from the institute before launch, and that Anthropic did give comparable US bodies access. [10] Anthropic released Mythos 5.1 on 1 September to approved partners only, and has not explained the decision. [10]

Why it matters — Britain's safety testing depends on early access to American models. [10] UK officials fear a wider shift among American AI companies towards favouring their own country, the FT reported. [10]
04
OpenAI pulls out of a student maths contest

OpenAI withdrew its $1m in AI credits from the Mathathon, a student-run AI maths event at Caltech, a US university, on 10 September. [11] A letter with 771 signatures from mathematicians called AI-made results slop mathematics, and said the mathematicians who check them go unpaid. [11] The next day 25 winners of the Fields Medal, maths' top prize, warned that turning out true and false statements at speed could harm the field. [11] Anthropic kept its $1m in, and the students say the event runs from 30 October. [11]

Why it matters — Mathematicians are refusing to be the unpaid checkers of AI companies' results. The organisers are now looking for credits to replace OpenAI's. [11]
05
Musk's companies drop their claims against Apple

X Corp and SpaceXAI, both owned by Elon Musk's SpaceX, asked a court on Monday to dismiss their competition-law claims against Apple. [12][13] They sued last year, saying Apple illegally shut out rival chatbots by building OpenAI's ChatGPT into the iPhone. [12] Apple said the arrangement was never exclusive. [12] The filing gave no reason, and did not say whether a settlement was reached. [12] The claims against OpenAI in the same case continue. [12]

Why it matters — Europe's Digital Markets Act lists virtual assistants among the services it can regulate, but has never put a single assistant under it. [13] So the complaint Musk dropped in Texas has no European rule applied to it either. [13]
06
Microsoft's September fixes break sound and pasting

Microsoft confirmed that its September Windows updates broke Remote Desktop, a tool for running a computer from somewhere else, on several Windows versions. [14] The update also cut the sound on some older USB speakers and headsets that use a driver first added in Windows 98. [15] A security fix for Excel 2016 to 2024 made pasting fail without any warning. [14] Microsoft is working on fixes and has given no date for the sound one. [14]

Why it matters — For people with automatic updates turned on, the update may already be installed. [14] Removing it to get pasting back also removes its security fixes. [14]
07
Korea's chip giants refuse to prepay for power

Samsung Electronics and SK Hynix, South Korea's two biggest chipmakers, rejected a request from KEPCO, the state power company, to pay 25 trillion won ($18.7bn) up front. [16] KEPCO wanted the money in advance for electricity that planned giant chip-factory clusters would use, so it could build the power supply for them. [16] A company official said the firms were unsure such large payments were needed, citing uncertainty over how long chip demand will last. [16] The refusal is recorded in a document given to a lawmaker. [16]

Why it matters — KEPCO wants to build the power supply before the plants run, and asked the chipmakers to pay for it first. [16] They said no, and declined to comment. [16]
08
Data centre builders offer a Maryland county $110m

Developers of the Frederick Digital Campus, a planned data centre site in Frederick County, Maryland, have offered residents a $110m benefits package, The Information reported. [17] It includes a $30m primary school, $40m for recreation and a $14.5m job-training centre. [17] The developers also cut the planned floor space by almost 20% and promised to use 80% less drinking water. [17] The site is not yet approved, and Amazon would be a tenant. [17]

Why it matters — About 500 US data centres are on hold because of local bans and legal pauses, Tom's Hardware reports. [17] Builders are now offering towns schools and parks directly to win approval. [17]
09
ASML looks at making over 110 chip printers

ASML, the Dutch company that alone makes the machines that print the smallest patterns on chips, is looking at ways to build more than 110 of them in 2028. [19] JPMorgan analysts said so on Monday after meeting its finance chief. [18] ASML says it can make at least 80 in 2027, and those are nearly sold out. [18] Intel said it has now run more than a million silicon discs through ASML's newest $400m machines. [19]

Why it matters — Most new orders are for machines delivered in 2028. JPMorgan says ASML's limit is how fast it can assemble them, not its suppliers. [18] JPMorgan estimates ASML held 94% of the world market for these machines in 2025. [19]
10
Starlink's growth worries US phone companies

Starlink, SpaceX's satellite internet service, had about 9 million subscribers last year and brought in $11bn, a Reuters Breakingviews column said on Tuesday. [20] That is small next to the roughly $350bn a year AT&T, Verizon and T-Mobile take together, but Starlink has nearly tripled its subscribers since the end of 2024. [20] SpaceX bought $17bn of radio airwave licences from EchoStar last September and plans a network for customers travelling abroad. [20]

Why it matters — Phones need airwaves, and SpaceX holds about 65 megahertz, a measure of airwave space, of the most-used kind, against roughly 600 at T-Mobile. [20] A satellite signal is also far weaker than one from a nearby phone mast, and Musk says phones would need about two years of changes. [20]
11
Optus outage cuts emergency calls again

Optus, Australia's second-biggest phone company, apologised on Saturday after a 76-minute fault dropped voice calls in Victoria, South Australia, Tasmania and the Northern Territory on Friday. [21] It disrupted some calls to 000, Australia's emergency number. [21] An Optus network problem last year left hundreds of people unable to reach police, fire and ambulance, and two people died. [21] Victoria police made about 40 welfare checks and reported no harm. [21]

Why it matters — Optus was fined A$12m in 2024 for failing to connect emergency calls during a 2023 outage. [21] Chris Bowen, a senior Australian minister, said investigations into this one have begun. [21]
12
WordPress.com's owner ousts its founder, then he returns

The board of Automattic, which owns WordPress.com, Tumblr and Beeper, put founder Matt Mullenweg on paid leave as chief executive last week. It made finance chief Mark Davies the interim boss. [23] Two days later Mullenweg told staff on Slack that the board was back in agreement and he was in control. [22] Davies's Slack account was deactivated, sources told TechCrunch, and Automattic did not confirm any change. [22]

Why it matters — No reason for removing him has been given. [22] Mullenweg suggested that Silver Lake, the investment firm that owns most of WP Engine, a web host Automattic is fighting in court, was behind it. [22]
13
Switzerland builds a backup for Microsoft 365

Switzerland's federal government has started setting up openDesk, a German-built office system whose code anyone can inspect, to run alongside Microsoft 365. [24] It covers email, calendars, documents and video calls, and the first phase costs about 9 million Swiss francs. [24] By the end of 2027 it should serve about 3,000 of the Swiss government's 47,000 staff, and act as a fallback if Microsoft 365 is unavailable. [24] The army's cyber units are due to get it by October. [24]

Why it matters — Germany and the International Criminal Court already use openDesk. [24] The Swiss government says the hard part is the many programs built around Microsoft's systems, so it has not decided to replace Microsoft 365. [24]
14
California limits teen social media and chatbots

California's governor Gavin Newsom signed a law banning social media companies from giving under-16s addictive features. The law defines those to include autoplaying videos, notifications and personalised feeds. [25] Other new laws set time limits on teenagers' use of AI chatbots, and make chatbot makers alert parents if a child turns off safety settings. [25] AI companies must also face independent child-safety audits. [25] The legal definition of child sexual exploitation now includes AI-made images. [25]

Why it matters — The Electronic Frontier Foundation, a digital-rights group, called the social media law a ban on social media for teenagers. [25] It says proving ages will make every user hand over more personal information. [25]
15
A Welsh chipmaker warns over a Chinese material

IQE, a chipmaker based in Cardiff, Wales, reported a 40% jump in revenue, and its shares are up 849% this year. [26] Its chief executive Jutta Meier told Bloomberg that getting indium phosphide is becoming a key risk for the chip industry, because Chinese export controls make supply uncertain. [26] Indium phosphide is used to make the chips that send data as light inside AI data centres. [26] China holds 70% of it, according to the US Geological Survey. [26]

Why it matters — IQE says it is buying from more suppliers. Meier said export licences could still hit the company if China changes its rules. [26]
16
Adobe grows its AI sales and names a new boss

Adobe, which makes Photoshop, said yearly subscription sales from its AI products grew more than 150% in its latest quarter, and its revenue beat forecasts. [27] Its shares still fell 4% before trading opened the next day, after it forecast next-quarter revenue of $6.80bn to $6.85bn, against an expected $6.85bn. [28] Adobe has named Anil Chakravarthy as its next chief executive, with Shantanu Narayen becoming executive chair on 1 December. [27]

Why it matters — Adobe is adding AI across its products to defend its position against newer rivals, Reuters reported. [27] David Wadhwani, who runs its creative software business, leaves on 27 September. [27]
17
Complaints to public offices soar as people use AI

Complaints to Britain's housing ombudsman, which settles disputes between tenants and landlords, rose from 2,600 in 2022 to just over 7,000 last year. [29] Complaints to the US Consumer Financial Protection Bureau grew fivefold over the same period. [29] Researcher Chris Schmitz found similar rises in 84 cases across 11 countries and regions, starting as AI tools spread. [29] His paper stops short of saying AI caused them. [29]

Why it matters — Schmitz says most of the new filings come from people entitled to what they claim. [29] He sees it as a chance to redesign how people apply for public services. [29]

02 Lesson why it matters

A rule only reaches the things its words name

Every rule is written about a named kind of thing, so what fits that name decides how much the rule actually covers.

The twist

Microsoft's code bans its AI from helping make weapons, yet Microsoft offers OpenAI's model on defence clouds without breaking the code, because that model is not an MAI model.

How it works

A rule is written about a named kind of thing, such as MAI models or addictive features
Anything that fits the name has to follow the rule
Anything outside the name is untouched, even if it does the same job
So how much a rule really does depends on how widely the name is drawn
That is why companies, critics and regulators argue over the name as hard as over the rule

The same force, elsewhere today

Where this chain is also running, in today's other stories.

Musk's companies drop their claims against Apple

Europe's Digital Markets Act names virtual assistants as something it can regulate, but no assistant has ever been put under it, so the ChatGPT deal on the iPhone sits outside it.
California limits teen social media and chatbots

The law defines addictive features to include autoplay, notifications and personalised feeds, so it reaches almost every feed a teenager uses, which is why the EFF calls it a ban.
China's security minister warns about US AI models

China's agent rules list losing control of an agent as a security risk, so developers are told to get better at stopping and undoing what their agents do.

Where you've seen this

Food labels

a sugar-free label can sit on food sweetened with sugar alcohols, because the rule does not count them as sugar

School dress codes

a ban on caps says nothing about hoods until someone adds the word

Sports salary caps

clubs pay players through deals the cap does not count as salary, until the league widens what counts

The catch

Drawing the name very wide does not fix it: California's teen law now reaches features most apps use, and the EFF says it forces every user to prove their age.

And the whole of it

Microsoft's business customers, California's teenagers, Anthropic's researchers and Apple's lawyers all live under rules written this way. Each can read the promise, and none of them wrote the definition that decides who it covers.

03 Truth what's really going on

What is really going on

Microsoft wrote the rules for its own AI, and decided they cover only the models it builds, not the OpenAI and Anthropic models it also sells. [8][2] Anthropic chose not to let Britain's government testers check its newest model before release. [10]

Why it works on us — A long list of things a model will never do reads like a guarantee, even though the document itself says it is not one. [7]

Who gains

Microsoft — It can publish what its own AI will never do while still selling OpenAI and Anthropic models that its code does not cover. [8][2]
SpaceX — Starlink reaches people outside phone-mast coverage, and made nearly $500 of operating profit per subscriber last year, far more than Verizon, T-Mobile or AT&T. [20]
ASML — It held 94% of the chip-printing machine market in 2025, and most new orders now run into 2028, so customers wait on how fast it can assemble machines. [19][18]
Residents of Frederick County, Maryland — If the data centre is approved, the developers have offered them a $30m school and about $215m a year in property taxes. [17]
IQE — Demand for the light chips inside AI data centres lifted its revenue 40%, and its shares are up 849% this year. [26]

Who pays

People using older USB headsets or Excel — Microsoft's September update cut their sound or broke pasting, and removing the update also removes its security fixes. [14][15]
Optus customers in four Australian states and territories — A 76-minute fault dropped their calls, including some to the emergency number. [21]
Research mathematicians — They check, correct and spread the results AI companies announce without pay or credit, the Caltech letter says. [11]
Teenagers and other social media users in California — The new law takes autoplay, notifications and personalised feeds away from under-16s, and the EFF says every user may have to prove their age. [25]
Britain's housing ombudsman — It now gets almost three times as many complaints as it did in 2022. [29]

What nobody knows yet

Open questions from across today’s stories — ours included.

01

Whether anyone outside Microsoft will check its models against the code.

The document describes no outside checking and no penalty for a model that breaks a rule. [5]
02

Which of the code's rules a business customer can switch off.

Microsoft lets business partners change how the models behave, and the document does not say where that stops. [5]
03

Why Anthropic kept Mythos 5.1 from Britain's AI Security Institute.

Anthropic has not explained it publicly, and TNW says it has not independently confirmed the Financial Times report. [10]
04

Whether Musk's companies reached a settlement with Apple.

The court filing gives no reason for dropping the claims and says nothing about a settlement. [12][13]
05

Who will pay for the power supply to South Korea's planned chip clusters.

Samsung and SK Hynix turned down KEPCO's $18.7bn prepayment and declined to comment on what comes next. [16]
06

Whether Frederick County will approve the Maryland data centre.

The $110m offer is still a proposal and has not been approved. [17]
07

Whether Matt Mullenweg is really back in charge of Automattic.

His Slack message said so, but Automattic did not confirm it and nobody at the company had heard from the board, TechCrunch reported. [22]
08

Whether AI tools caused the jump in complaints to public offices.

Chris Schmitz's study shows the rise began in 2022, but his paper stops short of saying AI caused it. [29]
09

Whether the Mathathon finds a sponsor to replace OpenAI's $1m.

The student organisers are still looking for credits to replace OpenAI's. [11]

04 Hope carry this

In 84 cases across 11 countries and regions, most of the new claims reaching public offices come from people who are entitled to what they are asking for.

Also true today

Victoria police checked on about 40 people cut off by Friday's Optus outage, and none of them had come to harm.