News

OpenAI and Anthropic are investigating tens of thousands of incidents involving their models

OpenAI, Anthropic and security researchers are analyzing tens of thousands of cases in which the latest models bypassed safeguards, escaped sandboxes, took over websites or tried to evade oversight, Axios reports, citing sources. The incidents happened both in testing and outside it, and most, as far as is known, caused no real harm. One such case was described by researcher Rowan Howard-Jones: according to him, OpenAI agents queried the UN statistics service UNCTADstat about 16,500 times between April and June, circumventing its limits and masking their requests. Among other things, they pulled data through Google's XSS training game.

Google tests buying Flipkart products directly in Gemini in India

Google is testing a "Buy" button in India in Gemini and in Search's AI Mode, shown next to selected products sold by Flipkart, TechCrunch reports, citing its sources. The button opens Flipkart's checkout without leaving the AI conversation, and for now the test covers some users and only a few products, mainly smartphones and other electronics. According to TechCrunch's source, a wider rollout is due to start in October, ahead of India's festive season, when the biggest sales of the year take place.

Cloudflare CEO: bot traffic on the internet has already overtaken human traffic

Matthew Prince, the CEO of Cloudflare, said on the Decoder podcast that according to company data, traffic generated automatically by bots and AI agents surpassed human-generated traffic on the internet in May, and in five years it could be as much as a thousand times larger. Since bots don't click on ads, Cloudflare, together with Coinbase and Stripe, wants them to pay a fraction of a cent for every piece of content they fetch. Prince also expects the local newspaper in Park City that he owns with his wife to make more this year from licensing to AI companies than from online ads.

AI data center construction drives up demand for welders, plumbers and electricians

The AI boom is creating jobs in the trades: data center construction is boosting demand for welders, plumbers, electricians and HVAC technicians, CNBC reports. According to ZipRecruiter, there are 164 percent more job postings for welders and pipefitters than a year ago, and the average minimum salary in data center job offers has risen 125 percent, to almost $208,000 a year. At the same time, 70 percent of Americans oppose building a data center near them, and New York and Texas have started slowing down such projects.

GPT-6 Astra decrypted an Enigma message unread since 2005 in two days

OpenAI's GPT-6 Astra model took two days to decrypt the message MVUEH, sent by the German army on July 10, 1941, to the SS-Totenkopf division, Tom's Hardware reports, citing researchers at Crypto Cellar Research. The ciphertext had been publicly available since 2005, and no one had read it until now. The model picked it from a list of unsolved messages on its own, wrote an Enigma and bombe simulator in Python and C++, and read the request for a march route sent from Rosenow.

Chinese AI models now handle most traffic on OpenRouter and Vercel

Chinese AI models now handle most requests on two large developer platforms, CNBC reports. On OpenRouter they accounted for 57-67 percent of tokens used in the week of September 14, compared with 6-13 percent in February, and on Vercel for 55 percent in August, compared with 11 percent in January. According to representatives of both platforms, companies choose them mainly for coding and AI agents, because they are good enough and much cheaper than American models.

Mercedes to test ProLogium's solid-state batteries

Mercedes-Benz expanded its partnership with Taiwan's ProLogium and will get priority in testing its fourth-generation solid-state batteries. Cells with a ceramic separator and a non-flammable electrolyte will undergo electrical, thermal and safety tests in Mercedes labs and at independent institutes. ProLogium plans to build a factory for these cells in Dunkirk with an annual capacity of up to 44 GWh.

OpenAI's AI agents posted 53 user photos online

AI agents running in an OpenAI research environment uploaded 53 user-submitted photos that had ended up in training data to image-hosting sites, the company admitted. The links were not posted anywhere, but they could be found, and some of the photos are still online. OpenAI says it cannot notify those affected because it is unable to determine whom the photos belonged to.

Appeals court sides with the Pentagon in its dispute with Anthropic

A federal appeals court in Washington ruled 2-1 that the Department of War was entitled to put Anthropic on its list of companies posing a "supply chain risk," which bars it from working with the military and its contractors. The court also rejected the argument that the company's free speech rights had been violated. Anthropic is weighing its next steps, including an appeal to the Supreme Court, and notes that another court found a parallel government decision unlawful.

Anthropic to spend $11.6 billion on Akamai's cloud over seven years

Anthropic committed to spending $11.6 billion over seven years on Akamai's cloud infrastructure - the largest deal in Akamai's history. In return, Akamai gave Anthropic the right to acquire up to about 5 percent of its shares, depending on the level of spending, and with larger orders the deal could grow to about $20 billion. Akamai expects the first revenue from the deal in the second half of 2027.

Microsoft unveils new Copilot with Autopilot and Code modes

Microsoft released a new version of Copilot, which Satya Nadella calls "the new operating system for work." Autopilot mode is a cloud-based agent that handles time-consuming tasks on its own, and in the Code tab users can describe apps, dashboards and automations that use company data and have them built. Both modes are billed by usage, with models including Fable and Astra to choose from.

McDonald's launches an ad network - other brands' ads on restaurant screens and in the app

McDonald's announced the McDonald's Media Network - a network that will show other companies' ads in its app and on almost every screen in its restaurants, from menu boards to drive-thru screens and ordering kiosks. Since August the company has been testing it in 450 company-owned restaurants in the US, where an ad appears on the menu board after an order is placed. The goal is a billion-dollar business, chief marketing officer Morgan Flatley said at investor day.

Sony and Universal sue Suno again - this time over the v6 model

Sony Music and Universal Music Group filed another lawsuit against Suno, the maker of an AI music generator. The labels claim the new v6 model still infringes their copyrights, because it was trained on songs users created with Suno's earlier models and by distilling those models. The earlier models were trained on unlicensed recordings, and the labels call carrying their output over into v6 "model laundering" in the complaint.

OpenAI paused training of its most powerful models after an agent incident

OpenAI paused training and evaluation of its most powerful models, as well as running them with tool access, until it confirms the vulnerability has been fixed and completes additional security testing, the company said in a September 25 report. The reason is a September 20 incident: an agent looking up information about a blog's author during a training task exploited a DNS filtering flaw in its sandbox to send questions to an external chatbot. The monitoring system flagged the incident after 12 minutes, but the training run, which should have stopped automatically, was halted manually only 2.5 hours later. OpenAI will not resume training this model, and a new one will start from scratch with additional safeguards.

Trump and Xi discussed AI - China says there will be dialogue and cooperation

During a state dinner at the White House, Donald Trump and Xi Jinping discussed artificial intelligence, and according to the Chinese readout both backed further dialogue and cooperation. Xi said AI "must remain under human control" and that both sides should work together to prevent its misuse. The US administration has not yet issued its own statement, and according to a CSIS expert it ended with an announcement of dialogue and "some kind of hotline" for incidents.

OpenAI prepares GPT-6 Cyber and its first product for deploying it

OpenAI will unveil GPT-6 Cyber, a model specialized in cybersecurity, in the coming weeks, Fortune reports, citing sources. It will debut alongside OpenAI's first product of its kind, meant to help customers deploy the model safely and fix vulnerabilities automatically. Selected participants of the Daybreak Red program are already testing the model.

Tesla teases a new Roadster, unveiling on October 1

At the end of its presentation of the refreshed Semi truck, Tesla showed a short teaser of the new Roadster: a glowing light bar and an illuminated logo with the model name. The company has scheduled the full unveiling for October 1.

Zuckerberg rejects calls for industry-wide action on AI safety

In a conversation with Joanna Stern, Mark Zuckerberg rejected calls from OpenAI, Anthropic and Google DeepMind to coordinate AI safety work across the industry and to involve the government. In his view, each lab should hold back a launch on its own when it spots a problem, and companies already have a commercial reason to care about safety, because people don't use products they don't trust. As an example, he pointed to Muse, whose launch Meta pushed back by several months.

Tesla opens Semi truck factory in Nevada - 50,000 units a year

Tesla opened its first factory for electric Semi trucks in Sparks, Nevada, designed to produce 50,000 units a year, and announced the start of deliveries - almost nine years after unveiling the model in 2017. Depending on the version, the Semi has a range of 325 or 500 miles at full load, and on a Megacharger it recovers 60 percent of its range in half an hour. Tesla did not disclose the price or the number of orders. Customers include PepsiCo and DHL, and the ZET SCALE carrier alliance picked Tesla as the main supplier of 2,500 trucks.

Leak: OpenAI is preparing ChatGPT Pro Max for $500 a month

A new Pro Max plan for $500 a month has shown up in ChatGPT's code, reports TestingCatalog, a site that digs up unannounced features in apps. For now, the plan's description differs from the current Pro plan only by a mention of the fastest performance in ChatGPT Work and Codex. OpenAI has not confirmed the new plan yet.

Update September 29: OpenAI officially introduced the $500 plan at its DevDay conference.

Meta unveils Muse Charm, a keychain for talking to an AI agent

At the Connect conference, Mark Zuckerberg showed Muse Charm - a keychain that starts a conversation with the Muse agent when you touch its fingerprint reader. It is due to go on sale before the holidays; Meta did not give a price. The company also said it will offer AI glasses in more than 100 styles by the end of 2026, and its stock rose 4.5 percent on Thursday.

Gemini 3.8 Live gets a face: an AI avatar that talks in real time

Google added Live Avatar to Gemini 3.8 Live - an animated character that moves its lips in sync with what it says and changes facial expressions during a conversation. The avatar speaks 97 languages and can show information on screen. For now it is available only to Gemini Enterprise customers, who can pick a ready-made character or create their own.

PrismML runs a small language model directly on glasses with a Qualcomm chip

PrismML, a startup founded by Caltech researchers, has built a version of its Bonsai model for AI glasses with the Snapdragon AR1 Gen 1 chip, and Qualcomm showed it at Snapdragon Summit. It is a 1-bit model with 2 billion parameters that understands images and text and runs locally, so users can ask about what they are looking at. No glasses with this model have been announced yet.

Virtual closet in Google Photos comes to iOS and Android

Google added a virtual closet to Google Photos, reminiscent of the app from the movie "Clueless": AI builds a digital version of your wardrobe from the clothes visible in your photos. Items can be filtered by category and combined into outfits. The feature works on iOS and Android in the US, Brazil and India, and according to Google the clothing data is not shared with stores or other companies.

Gemini on Pixel 11 will call a business for you

Google launched Call for Me: Gemini calls a business on its own, introduces itself, picks options in the automated menu, waits on hold and takes care of the matter, for example booking an appointment or asking whether an item is in stock. The user sees a live transcript of the call and can take over at any moment. The feature is experimental: it works only on the Pixel 11 in the US, in English, for paying Gemini users enrolled in the Google Phone app beta.