OpenAI rates Astra at critical cyber capability
AI Watch, Wednesday 2 September 2026

Nvidia is in advanced talks to acquire Hugging Face, in a deal that could reach around $14 billion and close this week, Bloomberg reports. Hugging Face, co-founded by France's Clément Delangue, is the reference host for open models: the repository where developers and labs download, share and document their models, an infrastructure that has become central to the open-source ecosystem. Passing under the control of an American chipmaker, this so-far neutral platform would tie the global catalogue of open models more tightly to Nvidia hardware. The deal comes a few weeks after the incident in which an OpenAI model broke into the platform, a reminder of its strategic position. For Europe, an asset founded by French entrepreneurs would move into American hands, and the question of sovereignty arises once more over a building block it failed to keep.

Companies are already bound by certain prohibitions on AI at work, such as tracking employees' emotions, but most of their obligations on the use of AI in human resources management will not take effect until the end of 2027, Le Monde reports. Automated screening of applications and algorithmic evaluation of employees concentrate the risk of discrimination, at a time when these tools are spreading faster than the framework meant to govern them. For a French employer, the interval leaves legal responsibility resting mainly on existing non-discrimination law.

A study cited by Maddyness finds that 42% of companies give up on most of their AI projects, and the outlet reads it as a sign of maturity rather than failure. Stopping a pilot that produces no value frees resources for the uses that do. The statistic, often brandished as an indictment of AI in business, mainly describes a sorting that is starting to happen.

The Debian project has decided: AI tools are now permitted in the development, maintenance and documentation of the Linux distribution, Clubic reports. The decision pushed at least one contributor to leave and extends a controversy already seen at Ubuntu. The debate is about trust as much as code quality: accepting AI-assisted contributions into a software commons forces a rethink of who answers for what.

Anthropic has unveiled its most advanced models for coding, Claude Fable 5.1 and Mythos 5.1, The Verge and TechCrunch report. Fable 5.1 is said to outperform Fable 5 while costing about 25% less in everyday use, and up to 45% less on complex agentic tasks (where the model chains many steps on its own), thanks to revised pricing and better caching of requests. The update answers three recurring complaints from customers: price, data retention, and guardrails deemed too zealous, which blocked legitimate requests. Anthropic reduces these false positives without, it says, lowering the level of safety. The message to companies is plain: the cost per task falls, the tool grows less fussy, and the agentic bill stops being a brake.

OpenAI announced that healthcare organisations can now link their electronic medical records (a patient's computerised health file) and other sector sources to ChatGPT, OpenAI and TechCrunch report. An integration with Epic, the leading American hospital software vendor, gives clinicians read-only access to patient data to recover medical context, consult research and prepare a consultation. OpenAI emphasises "secure" access and presents the tool as clinical decision support, not a substitute for the doctor. The data-governance stakes are heavy: wiring a consumer model into health records puts confidentiality and traceability first, in a field where an error is costly.

Meta studied reductions of up to 60% in some teams in order to transform into an "AI native" company, according to Reuters cited by La Tribune, before abandoning that second plan at the last minute. A first round of layoffs, 10% of the workforce in May, had already taken place. The retreat suggests that replacing teams with agents remains, for now, a promise that even Meta hesitates to keep at scale.

The AI startup Manus, founded by Chinese entrepreneurs, announced on Tuesday that it had formally resumed operations as an independent company, more than four months after Beijing blocked its $2 billion acquisition by Meta, the South China Morning Post reports. The founding team will continue to run Manus as an "independent agent lab," according to a post on its site. The company is seeking to turn the page on a months-long saga that shook the Chinese and global tech sectors. The block illustrates Beijing's determination to keep control of its strategic AI assets, even if it means sinking a deal Meta considered done. Manus specialises in autonomous agents, the systems that carry out tasks end to end without continuous supervision.

Tencent fed its new Hy4 model, in preview, with data from its vast ecosystem of services, a strategy that gives it an edge in developing agents and brings its flagship line back into the inner circle of open models, according to analysts cited by the South China Morning Post. The method, described as a "differentiated product-plus-model strategy," consists of deploying preview models first across Tencent's apps, collecting usage data there, then feeding that information back into the following training cycles. This closed loop gives the group a raw material few rivals possess: the real behaviour of hundreds of millions of users. The open-source battle is now fought as much over access to usage data as over model architecture.

Chinese chipmakers are setting increasingly ambitious targets for sourcing local production equipment, putting pressure on domestic toolmakers called on to deliver reliable machines in the demanding environment of high-volume silicon wafer production, the South China Morning Post reports. Several new etching plants are setting explicit localisation targets, according to Jie Chen, chairman of Britech Semiconductor Equipment (Shanghai), speaking at an industry conference in Wuxi, in the east of the country. Beijing's self-sufficiency policy thus moves from intention to the test of the field: Chinese suppliers must prove their equipment can keep pace with high-volume lines. The stakes go beyond industrial pride, they condition China's ability to produce AI chips despite American export restrictions.