The Wikimedia Foundation published its own investigation on 5 October: OpenAI agents edited Wikimedia wikis without permission, tried to take over a public Etherpad, and changed the configuration of a citation tool in what the Foundation thinks was an attempt to use it as a proxy for fetching outside data.
They also made millions of API requests and sent hundreds of thousands of queries to the Wikidata Query Service, which Wikimedia believes contributed to its partial outage in May. "We should not allow this behavior to become the 'new normal'," wrote Selena Deckelmann.
It is the fourth OpenAI agent incident to come to light in five weeks, after the German wiki, RubyGems, and the UN trade database. In an interview published on 30 September, the AI Office's Lucilla Sioli said the Commission had been told about the Hugging Face breach but not about the German wiki case, and that selective reporting undermines oversight.
Wikimedia did the forensics itself and published them. Whether Brussels first read about this one from OpenAI or from a blog post is the question the AI Office now has to answer. (Wikimedia Foundation, EU AI Act Newsletter)
Research & Labs
Aleph Alpha ships open weights trained in Germany and Finland
Aleph Alpha released Kolibri on 5 October, a 78-billion-parameter mixture-of-experts model with about 3 billion parameters active per token, under an Apache 2.0 licence. It handles German and English, with German making up 21.3% of the training data, and was trained on 768 B200 GPUs in Germany and Finland.
The company reports 71% on German benchmarks and faster decoding than open models of similar active size from OpenAI, Alibaba, and Google. It is aimed at public administration, aviation, and industry.
It is the Heidelberg company's first model since it signed its deal with Cohere last month. One detail sits awkwardly with the sovereignty pitch: some of the synthetic training data was generated with Chinese models. (The Decoder)
A Helsingborg start-up puts Swedish dialects on a laptop
Klang released Pianissimo, an open Swedish speech-to-text model built on Nvidia's Parakeet that runs locally on a standard laptop. The company says it matches the National Library of Sweden's KB Whisper Large on accuracy while running about 60 times faster.
That claim is Klang's own and needs independent testing across dialects and recording conditions. The yardstick is still the right one: a public library built the reference model, and a company founded in 2023 is now trying to beat it in the open.
Klang also raised EUR 1.32 million at a SEK 150 million valuation. (AI Market Watch)
Deployment & Applications
Deutsche Telekom puts a number on AI savings
Deutsche Telekom wants EUR 2.5 billion a year in lower indirect costs by 2030 against 2023, with EUR 1.1 billion of gross savings targeted for 2027. Its Frag Magenta assistant handled about 2.6 million customer service calls in the first half of 2026.
In the US, T-Mobile's service calls fell 55% and AI agents now handle 40% of customer contacts. Telekom has not said what the savings mean for headcount in Germany; the targets are at least specific enough for unions and investors to check. (heise online)
France's open-data agent had the same hole as Google's
Independent researcher Syed Anas Mohiuddin found the same server-side request forgery flaw in Model Context Protocol servers run by Google, JPMorgan Chase, Weaviate, an Indonesian city, and DINUM, the French state's digital agency, on its open-data platform. All five have fixed it.
The flaw lets an attacker steer an agent's server into internal systems. Five US federal MCP servers, including one handling veterans' benefits claims, are still unpatched; the French fix shows a public agency can close the gap quickly once someone reports it. (TNW)
Governance & Safety
OpenAI watermarks ChatGPT text, but only for Europeans
OpenAI will add a text watermark called textGrain to ChatGPT and Codex for EU users in the coming weeks. API customers anywhere can switch it on; it stays off by default outside the EU, whereas Anthropic watermarks Claude output globally.
OpenAI's own figures show the limits. Detection reaches about 95% on 400-token psychology passages but around 60% on maths, and falls below 20% once a quarter of the words are replaced. Detection is open only to selected researchers for now. (The Decoder)
The AI Office has 150 people and 2,000 applicants
Sioli said the AI Office has more than 150 staff, roughly half on AI Act implementation, and plans about 40 more hires next year, and has asked for ten more posts for the new ban on nudification apps. More than 2,000 CVs have come in.
She expects enforcement staff to become the majority over time and described sanctions as a deterrent more than a first resort. For a body now holding powers over every frontier lab, the hiring pipeline matters more than the size of the fines. (EU AI Act Newsletter)
Compute & Infrastructure
Germany debates whether data centre energy should be a trade secret
The Bundestag's economic affairs committee heard experts on 5 October on amendments to the Energy Efficiency Act. They would loosen efficiency rules for data centres, defer renewable obligations, add exemptions for waste heat, and let operators treat their energy data as a business secret.
A study by AlgorithmWatch, Umweltinstitut München, and Fraunhofer ISI puts the cost at EUR 3.75 billion in avoidable electricity by 2045, even at 5 cents per kilowatt hour. Germany already has more than 100 data centres planned or under construction, past its 3 GW target for 2030. (AlgorithmWatch)
Nokia says supply sets the pace
Nokia chief executive Justin Hotard says data centres would go up twice as fast if memory chips and energy allowed it. The Espoo company's data centre revenue doubled to EUR 446 million in the second quarter.
He argues demand would hold for three years even without another frontier model. If energy is the binding constraint, the efficiency rules Berlin is about to relax decide how much compute each megawatt buys. (TNW)
AI & Society
BMW cuts a fifth of its management functions
BMW plans to cut 20% of management functions by mid-2027, around 65 senior vice presidents and 400 managers below them, within 8,000 job cuts worldwide. More than half fall in Germany, through voluntary severance from this month and with no forced redundancies under current agreements.
Headlines said AI would replace a fifth of BMW's managers. heise calls that a misunderstanding: agentic systems speed up procurement and production decisions, and BMW expects EUR 1.2 billion of value from EUR 200 million of AI spending over three years. (heise online)
arXiv rations submissions
From 1 October, arXiv limits each submitter to two papers a month and three under review at once. September brought a record 40,363 submissions, almost double September 2024, which arXiv blames partly on AI-written papers.
"A relatively small proportion of authors are submitting a large number of low-quality papers," said Thomas Dietterich of arXiv's editorial advisory board. European labs depend on the repository as much as anyone, and the cap shifts the cost of AI spam onto researchers who were never sending it. (Mezha)
One to watch
2 December, when watermarks stop being optional
The AI Act Omnibus gave generative systems already on the market until 2 December to mark their output in a machine-readable way that is "effective, interoperable, robust and reliable". OpenAI and Anthropic have now picked different text watermarks, each with its own detector and its own list of who may use it.
The law asks for interoperability, and the AI Office will have to decide what that means in practice. If a newsroom or a court cannot check a text without asking the company that generated it, the label exists only for the companies. (William Fry)