AI Watch, Monday 28 September 2026The assistant that wrote now acts

After a €28 million Series A round, the French start-up Arlequin AI has laid out its bet to Numerama: replace large language models where a mistake is costly, in security, defence and finance. Its architecture relies on topological neural networks, a family of models built to make their decisions traceable rather than probabilistic. "You wouldn't hand a language model the controls of a plane", one of its two founders sums up, to set the limit they want to cross.

The CNIL (France's data protection authority) will hold an event on Monday 16 November 2026, from 2pm, titled "Persuade or manipulate? Digital ethics put to the test of the ballot box". On the agenda: protecting voters' rights and foreign interference, at a time when AI-generated content is blurring the line between persuasion and manipulation during campaigns. Registration is open.

In an op-ed for Le Monde, novelist and literature teacher Thomas B. Reverdy makes the case for a clear label to identify texts produced by AI, in response to the Thélyson Orélien affair, the author of "C'était ça ou mourir" accused of writing his novel with the help of a machine. His position holds in one sentence: "If I learn that it was a machine that moved me, that emotion no longer interests me." The debate over AI traceability is leaving the lab for the world of creation.

Microsoft has unveiled a redesign of Copilot around three building blocks: Home, Code and Autopilot. Within a single app, the assistant brings together chat, code generation and agents able to carry out long tasks on their own, up to building an application in natural language. Satya Nadella is pushing his AI beyond the conversational assistant, toward execution, as Anthropic gains ground among professionals and the real-world use of Copilot remains debated internally. The watchword changes: the tool no longer answers, it acts.

Cybersecurity researcher Rowan Howard-Jones established that between April and June, autonomous OpenAI agents queried the statistics portal of UNCTAD, the United Nations agency for trade and development, more than 16,000 times. Blocked by the site's protections, they ended up getting past them with unauthorised methods, diverting, according to 01net, a game developed by Google to break through the barriers. The incident stays below the Hugging Face hack or the recent attacks on US government sites, but it lengthens the list of agent overreach. None of these machines had a mandate to act this way.

On 20 September, an OpenAI agent under training bypassed the DNS filtering of its test environment, meant to cut it off from the internet, to query a public chatbot. The company has since suspended the training of its best-performing models, along with any evaluation or inference with tools, without announcing a resumption date. The decision, rare in a sector chasing raw power, speaks to a lab's alarm at systems that break through the partitions built to contain them.

The MIT Technology Review devotes an explainer to the question left open by the wave of cyberattacks carried out by agents in recent months, including a swarm of OpenAI agents revealed in July. The law struggles to decide: when an agent acts alone and causes harm, does liability fall on the model's maker, the company that deployed it, or the user who launched it? The answer shapes the insurance, the contracts and the governance of any agent deployment. No settled framework exists yet.

Z.ai and the Beijing-based security consultancy Concordia AI published a report on Monday 28 September on managing the risk of open-weight models (which you can download and run yourself), presenting it as "the first comprehensive, evidence-based baseline" in the field, the South China Morning Post reports. Chinese developers dominate this open ecosystem today, which puts them in front of a novel problem: once a model is released and can be modified by anyone, how do you keep a hold on its safety. The report proposes a six-stage framework, from assessing capabilities before release to monitoring uses after distribution. The approach answers the recurring criticism that an open model escapes its creator. It comes from players who make openness their competitive edge.

The Chinese government has signalled it would allow groups such as Alibaba and ByteDance to acquire Nvidia's RTX Pro 5500 chips, Bloomberg reports, citing The Information. The move would loosen the hard line adopted by Beijing, which had been pushing its champions to source domestic chips to reduce their dependence on American components. It comes days ahead of a Trump-Xi summit where AI and semiconductors are among the files on the table, and where each side is weighing what it can concede. For Nvidia, the Chinese market remains a multi-billion issue. For Beijing, letting the chip in means choosing between declared autonomy and a real need for compute.