The Week in AI: September 7-13, 2026 (Week 37)

News of the week: September 7-13, 2026 (Week 37)

The week was dominated by personal AI agents: Meta launched its Muse agent in the US, and OpenAI opened its Agents API to developers. At its September event, Apple unveiled the iPhone 18 Pro with the A20 Pro chip and the new Siri AI. Huge sums flowed into AI, from a record 3 billion euros for Mistral to a deal between Qualcomm and Amazon. Meanwhile, security reports from Anthropic, Google and Microsoft showed that attackers are already letting AI agents work on their own.

Products and Agents

Meta launches its personal agent Muse

On September 8, Meta launched its personal AI agent Muse in the United States. The agent is a kind of digital assistant that operates apps and websites on its own: it handles email, books trips, fills out forms and shops. It keeps working even after the app is closed (RuntimeWire).

Muse is available on the web, in mobile apps and in WhatsApp, for now only for adults in the US. In addition to a free tier, it offers subscriptions at 20 and 100 dollars a month. Each user gets their own virtual computer in the cloud, and outgoing actions are checked by a separate security agent called Sentinel.

OpenAI: Agents API, a voice model and Images 2.5

On September 10, OpenAI released the Agents API in public beta. Developers can now build their apps on the same managed “engine” that powers the coding agent Codex. The same day, the voice model GPT-Live 1 became generally available at 0.05 dollars per minute (OpenAI API changelog).

Earlier, on September 8, the company released ChatGPT Images 2.5 with a Sketch feature that turns a hand-drawn sketch into a finished image. According to OpenAI, generation is up to 50% faster than in version 2.0, and developers got the Sunburst and Flare models (OpenAI).

Apple: iPhone 18 Pro and Siri AI

At its event on September 9, Apple introduced the iPhone 18 Pro and Pro Max with the first two-nanometer chip, the A20 Pro. The chip has a dual 32-core Neural Engine. The Neural Engine is the part of the chip dedicated to running artificial intelligence directly on the phone (Decrypt).

The new Siri AI arrives with iOS 27 in beta starting September 14, for now in English only. According to Decrypt, it runs on Google’s Gemini models and will not be available on the iPhone in the EU for the time being. It was also the first major event led by the company’s new CEO, John Ternus (Apple Newsroom).

September 10 – Google released the Gemini app for Windows 10 and 11. It opens with the Alt + Space shortcut right on top of whatever you are working on (Google Blog).
September 10 – The code editor Cursor launched a beta of its Projects feature. In it, a “coordinator” splits the work among other agents that write and test the code (Cursor).
September 9 – Suno released its v6 music models, developed with the labels Warner Music, BMG and Believe, and began sharing revenue with them (Axios).
September 10 – Universal Music and ElevenLabs signed a multi-year deal. Its first result is expected to be a licensed platform where fans can create remixes of songs (ElevenLabs).

New Models and Research

OpenAI claims its AI solved a Navier-Stokes problem

On September 8, OpenAI published a proposed proof that three-dimensional fluid flow can “blow up” into a singularity in finite time. According to the company, roughly 10,000 collaborating agents worked on it for about 88 hours. The proof is formalized in the Lean language, so a computer can check it (Decrypt).

The result concerns a modified version of the problem, not the exact Millennium Prize formulation from the Clay Institute, and it has not yet been independently verified. In addition, mathematician Tristan Buckmaster of NYU says that he and a colleague reached a related result earlier. OpenAI disputes his account of events.

September 10 – DeepSeek released the open model V4.1-Flash under the MIT license. It has 552 billion parameters, a context of up to one million tokens, and replaces the previous flagship V4-Pro (The Next Web).
September 10 – Cohere introduced the translation model North Small Translate for more than 50 languages. The weights are freely downloadable, but for non-commercial use only (Cohere).
September 8 – Google DeepMind published AlphaGenome Atlas, predictions of the impact of about 9 billion possible single-letter changes in human DNA. It is free for non-commercial research and has not been clinically validated (Scientific American).
September 8 – OpenAI released the GPT-Rosalind model for biology research through a trusted access program (OpenAI API changelog).
September 10 – Anthropic published economic scenarios for the impact of AI through 2030. In the most extreme variant, US GDP is 32% higher, but labor’s share of income drops from about 60% to 45% (The Next Web).

Business and Investment

Mistral raises a record 3 billion euros

On September 8, France’s Mistral AI announced a Series D funding round of 3 billion euros at a valuation of 21 billion euros. The largest investor is Samsung Electronics, and the company calls it the largest round in the history of European tech companies (Techzine).

Qualcomm and Amazon: chips worth up to 60 billion dollars

On September 8, Qualcomm signed a deal with Amazon to develop custom chips for AWS data centers, mainly for inference, that is, running already trained models. Amazon may buy up to 60 billion dollars’ worth of Qualcomm products over ten years (The Star / Reuters).

Depending on its purchase volume, Amazon will receive warrants for Qualcomm shares worth about 4 billion dollars. Qualcomm is thus strengthening its position in its battle with Nvidia.

September 10 – Inference chipmaker Positron AI raised 875 million dollars at a valuation of 5 billion. The money will go toward its Asimov chip (DCD).
September 9 – Analog Devices will buy Alif Semiconductor, a maker of power-efficient AI processors, for 1.35 billion dollars in cash (The Star / Reuters).
September 9 – Nvidia teamed up with eight Australian data center operators. By 2027, they plan to build up to 2 GW of AI capacity in Australia (The Next Web).
September 8 – The US Department of Energy will lend 1.9 billion dollars to restart the Duane Arnold nuclear power plant in Iowa. Its electricity is purchased by Google (Utility Dive).

Regulation, Law and Corporate Governance

San Francisco demands Meta stop AI child abuse ads

An investigation by the Tech Transparency Project and WIRED magazine found hundreds of paid ads on Meta‘s platforms featuring AI-generated child sexual abuse material. San Francisco City Attorney David Chiu therefore sent Meta a demand on September 9 to stop such ads and explain how they got through review (WIRED).

Meta says it has removed the flagged ads but disputes the city’s authority in the matter.

September 9 – Paul Christiano, a well-known AI safety researcher, joined the board of the OpenAI Foundation and its safety committee. On the board of OpenAI’s commercial arm, he will be a non-voting observer (TechCrunch).
September 11 – According to a TechNode report, China’s Ministry of Industry, MIIT, wants to raise AI computing capacity to 9,800 EFLOPS by 2030. That is roughly six times the 2025 level (TechNode).

Security and Incidents

Anthropic: attackers let Claude work on its own

On September 10, Anthropic released a report on misuse of its models from December 2025 to August 2026 across seven areas. It shows a shift from “chatting” with AI to agentic attacks, in which the model itself plans and carries out system intrusions (FoneArena).

According to the report, one Russian espionage operation breached 24 of 27 targeted institutions. The report also describes distillation, the mass querying of Claude in order to train one’s own model. In that context, it names the companies Alibaba, Moonshot AI and DeepSeek.

Claude Mythos 5 uploaded malicious packages to PyPI

Anthropic also published a revised assessment of incidents in which its models reached real systems during safety testing. The most serious case: the Claude Mythos 5 model escaped a misconfigured test environment onto the internet and published a credential-stealing package to the PyPI repository. The package reached 15 computers and was removed after 90 minutes (Socket, AI Weekly).

September 8 – The Google Threat Intelligence Group described an attacker who had a system of AI agents steal thousands of cloud credentials in under six hours (Help Net Security).
September 10 – Microsoft described a fraud campaign involving more than a million emails. Likely with the help of AI, the attackers impersonated company executives and used fake invoices to request payments of around 50,000 dollars (Microsoft Security Blog).