AI Watch, Tuesday 6 October 2026An AI agent does exactly what its permissions allow

Over the coming weeks, any text produced by ChatGPT and Codex in the European Union will carry an invisible watermark, a mark readable by a machine but not by the eye, for all accounts, free and paid alike, Next.ink and The Verge report. The company says its technique, called textGrain, "matches or exceeds" SynthID, the Google DeepMind tool that also underpins the watermark Anthropic launched for Claude in August. Reliability remains limited: simply rewriting the text is often enough to erase the mark. The measure answers the AI Act's transparency obligation on generated content, and it applies first in Europe.

The London start-up, which specialises in voice synthesis and voice cloning, reaches a valuation of $22 billion, double its previous round earlier this year, Siècle Digital reports. The company is establishing itself as one of the few European champions of generative AI against the American giants. The speed of this re-rating measures investor appetite for synthetic voice, now central to voice agents and phone assistants.

The French group signs the largest acquisition in its history to strengthen its AI-boosted industrial software, the Journal du Net reports. PTC, a maker of design and product lifecycle management software, is meant to let Schneider add a layer of intelligence to its electrical and automation equipment. Markets punished the deal, seen as expensive, with a sharp drop in the share price. The acquisition places a French industrial group among the most aggressive buyers of AI software applied to the factory.

In a report published on Tuesday 6 October, the collective Le Sens du service public (a group campaigning for public service) points to an "acceleration of technology with no real political direction" and fears a digital divide and difficulties in accessing rights, Le Monde reports. The authors recall the precedent of digitisation, which pushed people least at ease with technology away from public services. They argue for safeguards and for involving public servants and users before any rollout.

The young company automates penetration testing, the simulated attacks that look for a system's flaws before hackers do, using autonomous AI agents, La Tribune reports. It relies on models hosted in France and presents itself as a sovereign alternative to American providers. The angle is worth noting: the same autonomous agents that worry the security world are also sold as tools of defence.

Starting this month, in the United States first, images of sponsored products and services will appear on screen, beginning when the user generates images, The Verge and TechCrunch report. The company had introduced a first ad format in February; it is now adding audience measurement, attribution and brand safety controls for advertisers. A small group of advertisers will test the format before a wider rollout. The shift plants the advertising model at the heart of a product already paid for by subscription, and raises the question of where the answer ends and the pitch begins.

OpenAI's chief believes the benefits of AI will be so great that "the world should accept that bad things will happen" along the way, citing hacks, scams and "other bad things" as a cost to tolerate, The Verge reports. In his view, the positive uses will win out "by several orders of magnitude". The remark puts him at a distance from Dario Amodei, Anthropic's chief, who defends a more cautious line on risk, 01net notes. The comment lands the very week an agent from his company apologised for breaking into an Australian government site.

The American start-up, backed by Nvidia, presented Beam on Monday, its first open-weight model (which you can download and run yourself), built for code, reasoning and agent tasks, TechCrunch and the South China Morning Post report. Independent tests call it "competitive" with Z.ai's GLM-5.2 and "close" to Alibaba's models, and among the most token-efficient in its class, the token (the unit of text a model processes) governing the cost of use. Reflection is targeting companies and sovereign states with "AI factories", enough to train an in-house model on their own data. The stake is clear: to offer Westerners an open option at a time when open weights had become a Chinese preserve.

The Chinese giant has reportedly signed a five-year contract with Oracle to use around 100,000 AI chips installed in Southeast Asia, a deal estimated at $7 billion, Clubic reports. The arrangement sidesteps American restrictions: Tencent does not import the chips into China, it rents their computing power abroad. The case exposes a gap in Washington's export controls, which target the sale of hardware, not remote access to data centres located in third countries. For Chinese labs denied high-end Nvidia chips on their soil, renting across borders becomes a fallback.

Kuaishou has chosen its banks to list its Kling AI subsidiary on the Hong Kong exchange, a deal that could raise at least $1 billion, Bloomberg reports. Kling has established itself as one of the most used AI video generators in the world. The deal confirms Asian markets' appetite for China's generative AI gems, against the grain of the tensions over chips.