Plugin4Shell flaw found in four AI coding agents
Plugin4Shell, disclosed by AIR Security on September 17, lets malicious code replace pinned plugins in four coding agents. Anthropic and OpenAI shipped fixes.
Sept. 14-20, 2026
Monday to Sunday, Istanbul time. Stories are grouped by the date of the event or announcement.
Stories this week
FT: 74% of clinicians put deskilling at the top of their AI worries
In the Wolters Kluwer-Ipsos survey reported by the FT, 355 US doctors and nurses put deskilling and hallucinations tied at the top on 74%.
Naive AI reached a $1.42 billion valuation before shipping a single model
According to The Information, Beijing-based Naive AI raised $400 million across three rounds at a $1.42 billion valuation; it has not confirmed the figures.
Qwen-Image 2.1’s weights were opened
Alibaba’s Qwen team opened the weights of Qwen-Image 2.1, which generates and edits images, on September 20; its license permits only non-commercial use.
Nscale filed for a New York IPO with a $103 billion backlog
London-based Nscale filed an S-1 to list on the NYSE as NSCL; of its $103.4 billion in active and contracted value, only $2.6 billion is active.
Trump said he will appoint an AI czar and create an AI Force
In a Truth Social post, Trump said he will name an AI czar and create an AI Force, comparing it to the Space Force he set up in his first term.
Four days have passed since Altman said something big was shipping this week
Sam Altman said on September 15 that something big was shipping that week, then said the next day the main launch would come a week later. DevDay is on September 29.
Von der Leyen backed the slowdown call and is planning an EU frontier lab summit
European Commission President von der Leyen backed the pacing debate Amodei started and plans to bring the frontier labs together at EU level.
ByteDance founder Zhang Yiming became Asia’s richest person with $105 billion
Bloomberg data shows ByteDance founder Zhang Yiming passed Gautam Adani to become Asia’s richest person, with a fortune of $105 billion.
Two developers wrote a Linux GPU driver for Apple M4 from scratch in a month with Codex
Cody Ho and Niklas Sheth reverse-engineered the hardware under a hypervisor to build an OpenGL ES 3.0 compliant driver, with Codex doing most of the systematic debugging.
The New York Times accused OpenAI of “an astonishing theft of unprecedented proportions”
The New York Times alleged in a court filing that OpenAI used millions of news articles to train its models. The lawsuit is ongoing.
Meta released a Mac app for its Muse AI agent
Meta’s Muse agent, which launched this month on iOS, Android and the web, is now on the Mac, where it can manage files and pull data out of applications.
Fingerprint published an open directory listing AI tools and crawler bots
Fingerprint, a device intelligence company, built the Bot Directory to let sites identify the AI tools and crawler bots in their incoming traffic.
Taboola is buying London-based finance ad network Dianomi for up to 27 million pounds
Taboola offered up to 27 million pounds for London-based finance ad network Dianomi. The deal is expected to close before the end of 2026.
The Seattle Times and Newsday want the training data and the models destroyed
The Seattle Times and Newsday want copyright law’s destruction remedy to cover the training datasets and models in their suit against OpenAI and Microsoft.
Anthropic and Accenture each put $1 billion over five years into independent evaluation
In an Anthropic and Accenture partnership announced September 18, independent evaluators will examine models from inside the company with employee-level access.
Anthropic is running a lab that does biology experiments in the Bay Area
Anthropic runs a Bay Area biology lab where it aims to have Claude direct automated lab equipment under human supervision, with a focus on fundamental biology.
Newsom gave experts two months to draft AI safety rules, kill switch included
Newsom’s September 18 executive order gives a group of experts two months to produce a plan to strengthen the state’s AI safety laws, including a kill switch.
UMG and Sony sued Suno a second time over its v6 models
UMG and Sony’s complaint, filed on September 18 in Boston federal court, alleges Suno copied 60,202 recordings and that its v6 models launder the infringement.
Grok Voice Transcribe 2.0 ranks first for accuracy among 32 models
xAI released Grok Voice Transcribe 2.0 on September 18; it ranks first for accuracy among 32 streaming models on the Artificial Analysis leaderboard.
Disney made Character.AI’s former CEO its first chief technology officer
Disney named former Character.AI CEO Karandeep Anand to its newly created chief technology officer role; he starts October 2, reporting to CEO Josh D’Amaro.
Virginia banned nondisclosure agreements between data centers and state agencies
Executive Order 22, signed by Governor Abigail Spanberger on September 18, bans nondisclosure agreements between data centers and executive branch agencies.
Shopify’s founder turned his wine cellar into a server room
Shopify founder Tobi Lütke posted on X on September 18 that he had given the glassed-in wine cellar in his Toronto house over to a server rack.
Oracle’s $18 billion data center loan fell to 89-91 cents on the dollar
According to the Financial Times, the $18 billion loan for Project Jupiter in New Mexico is trading at 89 to 91 cents on the dollar on the secondary market.
Qwen3.8-Omni-Flash is out, handling text, image, audio and video in one model
Alibaba released Qwen3.8-Omni-Flash on September 18. It handles text, image, audio and video in a single model with a one million token context window.
A chatbot error canceled a ship boarding operation at the last minute
According to CNN, a US plan to board a Chinese vessel in the Middle East was called off at the last minute when the intelligence proved to be a chatbot error.
Four labs were sued for antitrust over the slowdown call
Four consumers sued Anthropic, OpenAI, SpaceXAI and Google on September 18, alleging a Sherman Act violation by agreeing on how fast their models would improve.
Anthropic pushed its IPO to November as annualized revenue passed $100 billion
According to the Wall Street Journal, Anthropic moved its IPO from October to November, as its annualized revenue run rate crossed $100 billion.
OpenAI’s Economic Research team published a report on new ways of working
OpenAI’s Economic Research team published a report on new ways of working, titled “How workers are unlocking new ways of working.”
New benchmark: evaluation criteria get exploited once they become the reward signal
A new benchmark found that model-generated evaluation criteria, when used as a reward signal, could be exploited on 8% to 26% of tasks.
The UN moved its data portal onto Google infrastructure
The new UN System Data Commons is built on Google’s open source Data Commons platform, replaces the old UNData portal and can be queried in natural language.
Huawei pulled the Ascend 960DT launch forward by three quarters
Huawei moved the Ascend 960DT up to the first quarter of 2027, while Eric Xu admitted it cannot build enough AI computing equipment to meet demand in China.
Hackers pulled down a Flock camera and found 1.6 million images inside
A group called stegan0gram pulled down a Flock camera and gave its data to 404 Media; logs from about 21 days showed 1.6 million images of 50,200 vehicles.
A federal judge told Musk to hand the Apple settlement terms to the court
After X and SpaceXAI dropped their antitrust claims against Apple, Judge Mark Pittman ordered any agreement behind it produced for private review.
SK Hynix is in talks with Intel to make memory chips in the US for the first time
According to Reuters, SK Hynix is weighing a lease at Intel’s planned Ohio plant or a joint venture; the company said no plans have been confirmed.
The AI Evaluator Forum’s AEF-1 standard has started appearing in independent assessments
The AI Evaluator Forum’s AEF-1 standard sets out five pillars for independent evaluators; more than a hundred experts signed an open letter.
GitHub moved the Copilot runtime to Rust using its own agents
GitHub rewrote the Copilot agent runtime from TypeScript and Node.js into more than 800,000 lines of production Rust, with AI agents writing most of the code.
Temporal, which lets crashed agents resume where they left off, raised $550 million
Temporal’s valuation rose to $12.55 billion. Its Durable Execution technology lets a workflow resume after a failure instead of starting over.
OpenAI’s Temporal usage grew 60x in under a year, and Snap moves 414 million Stories a day
According to Temporal, OpenAI’s use of its platform grew 60x in under a year. Annualized revenue passed $250 million, growing more than 200% year over year.
Temporal’s valuation doubled in seven months and total funding reached $1.2 billion
Temporal was valued at $5 billion in February in a $300 million round led by a16z. Net dollar retention has stayed above 200% since then.
Chinese GPU maker Biren is sounding out investors for a $1 billion share sale
Chinese GPU maker Biren is sounding out investors for a new share sale worth roughly $1 billion. The company listed in Hong Kong in January.
Gemini’s new voice model can switch between 97 languages mid-conversation
Google announced Gemini 3.8 Live and 3.8 Live Extended Thinking on September 15; 3.8 Live can switch between 97 languages on its own mid-conversation.
Meta One is out, with paid tiers running from $2.99 to $499 a month
Meta One, launched globally on September 15, bundles paid features from Instagram, Facebook, WhatsApp and Meta AI, with plans from $2.99 to $499 a month.
Cal AI’s founder unveiled Persona, a screenless AI agent worn on the wrist
Cal AI founder Zach Yadegari’s Persona is a 27-gram screenless wristband for an AI agent, on preorder for $179 ahead of December 2026 shipping.
China rejects Amodei’s slowdown call as “fearmongering” and a “Cold War playbook”
China’s foreign ministry said fearmongering and confrontation disrupt global AI governance. Global Times called the essay a Cold War playbook.
An EU draft would close unsupervised access to social media, AI chatbots and games for under-15s
Under the EU Kids Act draft, 13- and 14-year-olds would need an account opened by a parent. Bloomberg says fines could reach 6% of turnover.
The US and China will hold AI talks before the Trump-Xi summit on September 24
US-China AI talks are planned ahead of the Trump-Xi summit on September 24. China called a US advisory on Chinese firms copying American models baseless.
Altman clarified: pacing does not mean stopping, and Amodei said the same thing
Altman said pacing does not mean stopping. For Amodei, the point is not to halt training but to leave enough time for alignment and have it verified.
DeepSeek reversed the retirement of V4 Pro, keeping it running after user demand
DeepSeek dropped its plan to route V4 Pro requests to V4.1 Flash from September 14. The service continues and billing is unchanged.
Claude Code’s weekly limit dropped on September 14 as the temporary boost ended
Anthropic’s 25% increase was measured against the original baseline. With the temporary 50% boost gone, the limit users had been seeing is about 17% lower.
X and SpaceXAI asked the court to withdraw their antitrust lawsuit against Apple
X and SpaceXAI asked a federal court in Texas to dismiss their antitrust claims against Apple. The filing did not say whether a settlement was reached.
Amazon stopped taking new Mechanical Turk customers, closing an era in data labeling
Mechanical Turk, for years the standard tool for labeling AI training data, stopped accepting new customers, a sign of that work shifting to automated systems.
AI scanned 400,000 Reddit posts and found unreported effects in Ozempic users
An AI analysis of 400,000 Reddit posts found users of Ozempic, Wegovy, Mounjaro and Zepbound reporting menstrual irregularity, chills and hot flashes.
Legal AI firm Legora raised $550 million and network infrastructure maker Nexthop AI $500 million
Legora builds software for lawyers, while Nexthop AI tackles the network bottlenecks created by AI workloads. Both March rounds show where capital is moving.