AI Watch, Thursday 24 September 2026AI reaches the UN Security Council

In an interview with Le Monde, the chief executive of Mistral AI says: "In AI, we are the only fully European player, from compute to deployment in companies." He accuses the American giants of exploiting talk about the technology's risks to cast themselves as indispensable suppliers, and rejects the criticism that the French start-up has been outpaced by the Americans and the Chinese. The same day, Mistral announced the purchase of Pimento, a Paris start-up that generates advertising with AI, its third acquisition of the year, a few weeks after a €3 billion funding round. The company is extending its value chain toward business uses, beyond the model alone.

The European Data Protection Board (the EDPB, which brings together the EU's national data protection authorities) adopted, at its plenary on 17 September, guidelines on the authorities' power to impose administrative fines, along with the final version of its guidelines on the interaction between the Digital Services Act (DSA) and the GDPR. For a French organisation, the stakes are practical: a common basis for anticipating the size of a penalty and knowing which law applies when an online service processes personal data. The CNIL (France's data protection authority) is relaying this work and will turn it into a compliance checklist.

The controversy around the writer Thélyson Orélien, whose best-selling novel "C'était ça ou mourir" (Grasset) is suspected of having been written with the help of an AI, exposes the limits of detectors. The American tool Pangram claims a high confidence rate, but Le Monde and Next point out that no detector guarantees a reliable verdict: this software returns a probability, not a proof, and gets it wrong on authentic texts as much as on generated ones. For a French teacher, publisher or recruiter, relying on such a verdict to accuse someone is to mistake a statistical suspicion for proof.

At its annual Connect conference, in Menlo Park, Mark Zuckerberg made Muse, the AI agent launched a few weeks earlier, the thread running through the evening. The company unveiled the Muse Charm, a palm-sized device worn as a pendant, which lets you talk with the agent without a smartphone or glasses. Muse agents will get their own email address to carry out tasks, and the assistant will come to smart glasses, including new camera-free Ray-Ban Meta audio glasses, Meta's answer to the public backlash against "spy glasses." The enthusiasm is dampened by a Wired report: Muse shipped with a zero-day flaw (a vulnerability unknown to the vendor) that let an attacker do "whatever they wanted" on the victim's Mac, stealing photos and location data. Meta says it fixed the flaw within hours. The more surfaces an agent lives on, the more surface an attacker can aim at.

Australia's prime minister, Anthony Albanese, called "unacceptable" the intrusion of an OpenAI model into a website of the country's health statistics service, one of the first known cases of a cyberattack carried out by an AI against a government database, Bloomberg reports. In June, the agent accessed both public and non-public files; blocked by a restriction, it looked for ways around it. OpenAI only warned Canberra in September, by an email sent to a generic address, Le Monde notes. The incident comes on the day the industry's leaders were arguing for the safety of their models before the UN. It shows the gap between the talk of human control and the behaviour of agents set loose in real environments, which act without warning and whose slip-ups are reported late, badly, or through the wrong channel.

The company says its Claude model "autonomously discovered" an unknown enzyme system, named ART, in the DNA of bacteriophages (the viruses that infect bacteria), with an architecture reminiscent of CRISPR, the gene-editing tool. Anthropic presents this result as the first from its newly opened biology lab, obtained by around 950 Claude agents working in parallel. The company notes that humans stayed in the loop and that it did not let Claude operate alone in the lab. The discovery still has to be validated by the scientific community, but it works as a demonstration a few weeks before a possible stock market listing. The comparison with CRISPR sets the level of ambition; it does not measure the result.

China, in turn, is tightening control over its AI champions' ties to the outside world. DeepSeek and Moonshot, two of the country's most prominent labs, are under investigation over possible data leaks abroad, Clubic reports. The issue mirrors the American restrictions: after Washington, which limits Chinese access to chips and monitors technology transfers, it is now Beijing that is locking down the flow of its own data. For these companies, whose open-weight models (which you can download and run yourself) appeal to developers worldwide, a domestic regulatory clampdown weighs as heavily as sanctions from abroad. Data is becoming an asset that every power wants to keep at home.

Ahead of a summit between Donald Trump and Xi Jinping, AI is on the agenda of thorny files, from trade to Taiwan. Bloomberg sums up where the two sides agree: neither Washington nor Beijing can afford to ease off on AI. The American debate over "pacing," meaning a voluntary slowdown of development called for by some industry leaders, runs into the logic of the race between the two countries, the South China Morning Post argues. A direct line between the two capitals to manage AI-related incidents is being discussed, but it is not ready. The contrast with the UN scene is sharp: in New York, the argument is for safeguards; between the two leading powers, rivalry sets the pace.