AI Watch, Friday 11 September 2026AI safety moves from the activist fringe to mainstream debate: Sam Altman says he is ready to slow frontier AI development and is testing whether such a brake would be legal, a former Anthropic researcher warns that \"AI could kill us all,\" OpenAI adds a worried researcher to its board, and Anthropic publishes 161 pages of Claude misuse, from Yemeni missiles to fake dating profiles. In Europe, the CNIL sets rules for electronic invoicing and the EU's cybersecurity agency trials Anthropic's most offensive model. In China, DeepSeek reignites the price war with its V4.1 Flash model and chipmaker Enflame jumps 188% on its stock market debut. Microsoft plans 38 gigawatts of data centres and Jensen Huang promises Nvidia 70% growth.

Since 1 September 2026, the reform of business-to-business electronic invoicing has come into force, requiring every company to issue and receive its invoices in digital form. The Commission nationale de l'informatique et des libertés (France's data protection authority) has published guidance to help companies secure the personal data that moves through these flows, in particular via the approved platforms that centralise exchanges. For finance and legal departments, the technical compliance project comes with a GDPR project that is best opened now.

The French open-model platform has added to its security.txt file, usually reserved for security researchers, a note addressed to the autonomous agents that crawl its site: no need to try to hack it, the answers to their goals are elsewhere. The wink reflects a new reality for security teams, namely that a site's visitors are no longer only human, and that agents launched by third parties now test defences around the clock.

ENISA, the European Union's cybersecurity agency, has been granted access to Mythos, Anthropic's most offensive model for cyber defence, Next.ink reports. In May, OpenAI had already opened up its GPT-5.5-Cyber to the agency. A European public institution is thus equipping itself with the same tools whose misuse Anthropic documents elsewhere, in the hope of strengthening its own defence before attackers make use of them.

The French start-up Ed.ai has closed a €5 million funding round to develop its grading-support tool, which analyses students' work to produce more precise feedback for teachers. The project aims to lighten the marking load while keeping the teacher at the centre of assessment. French edtech finds here a concrete use case, where AI prepares the work without replacing the teacher's judgment.

Sam Altman told his teams that OpenAI was considering slowing the development of its most advanced models, and that he hoped to see other labs do the same, according to Bloomberg. At the same time, the company is trying to find out whether industry coordination to slow down might fall foul of competition law, Wired reports. The shift comes with a change of mood: Jacob Coxon, who has just left Anthropic, warns that "AI could kill us all" in the years ahead, a message picked up by Platformer, Wired and Bloomberg. OpenAI has also added to its foundation's board a researcher known for pessimism about AI risks. The existential question, long confined to a few circles in the San Francisco Bay Area, is settling at the centre of the sector's conversations.

On 10 September, OpenAI unveiled a version of ChatGPT dedicated to finance, combining built-in financial data with its latest model, GPT-6 Astra, for research, modelling and the production of client-facing documents. The tool is aimed at analysts, bankers and portfolio managers, promising to speed up market analysis and the preparation of materials. The announcement extends OpenAI's strategy of tailoring ChatGPT by sector, with the parallel launch of a "Data agent" able to connect a company's data and turn it into dashboards. Finance, a heavy consumer of document synthesis, is becoming a leading commercial target.

Short of capacity, Microsoft has had to turn down some AI and cloud orders, Bloomberg reveals. The group now plans to add 26 gigawatts of computing power and bring its fleet to 38 gigawatts of total capacity, the equivalent of some thirty nuclear reactors devoted to its data centres alone. The plan illustrates the current bottleneck, which owes less to chips than to electricity and available land. The AI race is now being run as much among power utilities as among model builders.

Nvidia's chief executive, Jensen Huang, said his company would grow 70% next year, while rejecting the idea that its deals with customers are "circular," meaning that it funds those who buy its chips, according to TechCrunch. At the Goldman Sachs technology conference, he named cybersecurity as AI's next big market, after coding, citing defence systems running around the clock. Nvidia is already forming partnerships in the field, Clubic reports. The supplier of the chips that train the models is signalling where it intends to sell the next wave.

On Thursday 10 September, DeepSeek launched its V4.1 Flash model, which it presents as more capable than its previous flagship while cutting the cost of inference (the cost of running a model once it is trained) and gaining speed. The model rests on a new architecture called "Causal-Encoder-Decoder" and on 552 billion parameters, with a "mixture of experts" system (only a fraction of the network activates on each query, which limits computation), whereas a conventional model draws on the whole network. DeepSeek claims V4.1 Flash beats Moonshot AI's Kimi K3 on cybersecurity and coding benchmarks. The announcement fits China's strategy of aggressive pricing and comparable performance, which is pushing down global AI prices.

Enflame Technology, one of Nvidia's main Chinese rivals, saw its share price climb 188% on Friday 11 September in its first trading session in Shanghai. The stock opened at 410 yuan, against an offer price of 142.18 yuan, taking the valuation of this Tencent-backed AI chip designer to 176.4 billion yuan, around $25 billion. Enflame was the last of China's four big AI chip champions to go public. The enthusiastic investor reception confirms the domestic market's appetite for local alternatives to Nvidia, as US export controls push China to build its own supply chain.