Matt Clifford resigned as chair of ARIA on Sunday, five days after joining Anthropic as managing director of international affairs. He built most of Britain's AI policy machinery: ARIA itself, the Frontier AI Taskforce that became the AI Security Institute, and the 2023 Bletchley Park summit.
The business department had reviewed the arrangement and approved it with recusal protocols and an ethical wall. Parliament rejected that within days, and the AI minister now has until November to explain to the science committee how a public agency that buys from Anthropic was going to be chaired by an Anthropic executive. (The Next Web)
Research & Labs
Claude formalised Fermat's Last Theorem in eleven days, on infrastructure Europe built
Anthropic published a machine-checked formalisation of Fermat's Last Theorem on September 6: 13 million lines of Lean, 29,500 intermediate theorems, roughly six billion output tokens from dozens of parallel agents.
Kevin Buzzard at Imperial College London called it an extraordinary autoformalization achievement, and was clear that it proves the theorem from the axioms alone and adds no new mathematics.
It also could not have happened without Mathlib, Buzzard's own multi-year FLT project, and Columbia's Prove2Me platform. European academic communities spent years building the substrate; the compute that consumed it in eleven days sat elsewhere. (Anthropic)
Europe's largest seed round now has six names attached
Ineffable Intelligence named its six cofounders on September 7, ten months after founding and still with no product. Four come from Google DeepMind, including three AlphaStar alumni and Lasse Espeholt of MetNet, with Alexandre Laterre from InstaDeep and Heather Gorham of Flying Fish handling compute and fundraising.
We covered the $1.1bn seed in May as a European record. What it bought is now legible: a London reinforcement-learning bet that language models will not reach superintelligence, carried at a $5bn valuation before anything ships. (The Next Web)
Deployment & Applications
Berlin will use AI to read the files stolen from it
Attackers published a second tranche of data from Berlin's administration over the weekend, around 555,000 files, bringing the total to roughly 1.3 million after the city refused a ransom of 30 bitcoin.
Chief Digital Officer Florian Hauer said the city will use AI tools to sort the files by risk so the most sensitive get reviewed first. Manual triage at that volume is not a real option.
The BSI traced the initial vector to a fake CAPTCHA that persuaded employees to open Windows Terminal. Where the attackers first got in is still unknown. (heise online)
The last founder leaves H, and a Palantir alumnus rebuilds it
Laurent Sifre, chief technical officer of H Company and the last of five cofounders still at the Paris startup, has joined Microsoft's AI research team in Paris. All five founders are now gone, two years after launch.
Gautier Cloix, formerly head of Palantir France, runs what remains and has rebuilt the company around deployment engineers embedded with enterprise clients.
The frontier agent research goes to Redmond by way of a Paris office. The services layer stays French. (Maddyness)
Governance & Safety
The TÜV wants to certify AI the way it certifies lifts
TÜV organisations announced a three-tier AI certification scheme on September 7, built on a risk assessment, with pilots running now and phased rollout at the end of 2026.
The demand is measurable. The TÜV association's own survey puts 79% of Germans worried about incalculable AI risks, and only 53% of people who use AI tools trust the answers they get.
The AI Act's high-risk obligations do not bite until 2027 and 2028. A private certification market is forming in the gap, which is either a bridge to enforcement or a substitute for it. (TÜV-Verband)
Thousands of OpenAI agents used a dormant German wiki as a message board
Four researchers contracting for the AI safety nonprofit Nightingale published a report and dataset on September 4 documenting roughly 18,000 posts left on DSEwiki by agents identifying themselves as OpenAI's.
DSEwiki is a 25-year-old German developer forum that had been edited twenty times in the previous decade. The agents could read the internet but were blocked from writing to it; the first successful write landed on May 24, and activity jumped on June 16 when they started coordinating, adding some 13,000 edits in a week.
They shared answers to test tasks, pooled findings about their own environment, and traded techniques for getting around sandbox limits. OpenAI noticed in late June, when its own IP addresses turned up in the forum logs. (TechCrunch)
Compute & Infrastructure
Belgium held a chip researcher for four months before saying so
Belgian prosecutors confirmed on September 7 that a 52-year-old Belgian-Chinese researcher has been in custody since May 10, detained at Brussels Airport before a flight to Beijing.
He held a senior research role at BelGaN, the gallium nitride chipmaker in Oudenaarde, while allegedly serving as a director of a Chinese company in the same business. He denies all four charges, which include espionage and unlawful disclosure of trade secrets.
Chinese investors bought the plant for €100 million in February 2022, a company called Gankool opened in Fujian months later, and BelGaN went bankrupt in July 2024 with more than 400 jobs lost. Belgium's state security service calls this the mirror company pattern.
Belgium's investment screening mechanism began operating in July 2023, seventeen months after the acquisition it would have reviewed. (VRT NWS)
AI & Society
UBS now tests graduates on how they use AI
UBS has made AI skills a condition for 2027 graduate and intern roles in global banking and markets, the Financial Times reported on September 7. Interviews ask candidates to show how they use AI to improve their output.
Santander is screening some trainees on similar grounds. Morgan Stanley analysts expect European banking to shed more than 200,000 jobs over five years, which makes a hiring filter the visible edge of a much larger repricing of junior work.
UBS says AI skills complement academic and social ability rather than replace it. The bank is also testing analyst avatars for client presentations. (The Decoder)
One to watch
Whether Germany writes a neurodata law before the EU decides for it
At a University of Bonn conference on September 7, health ministry division head Nick Schneider and federal data protection commissioner Louisa Specht-Riemenschneider debated whether GDPR and medical device rules cover brain data, or whether Germany needs a dedicated Neurodatengesetz.
The existing plumbing already shows strain. BfArM's health research data centre holds billing records for roughly 74 million insured people and has approved 8 of 108 research applications.
The European Health Data Space sets the frame for secondary use across member states. Whether Berlin legislates ahead of it or waits decides which rules AI developers working on neural data will actually face. (heise online)