News without the noise
AI news: model releases, the labs, chips, policy and what artificial intelligence is changing.
847 articles · page 8 of 18
Three former OpenAI safety researchers say they were fired after raising concerns that the company was not doing enough to monitor the behaviour of AI agents, while OpenAI says they were let go for breaching company policy by mishandling information.
Anthropic and OpenAI have hired at least five former Trump administration officials this year, placing them in senior roles as the AI industry works to strengthen its relationship with Washington and navigate growing national security concerns around frontier technology.
· 2 sourcesAmazon Web Services has released Strands Decider 2B, an open-weight decision model built on Alibaba's Qwen3.5, adopting the same API endpoint — /v1/systemone — that startup TypeSafe AI introduced for its Jev model just weeks earlier. The move signals rapid standardization around a new category of lightweight models designed to make fast, probabilistic decisions for AI agents, without generating prose or code.
Microsoft will let Microsoft 365 Family and Premium subscribers share Copilot and other AI benefits with up to five other people on their plan, ending the current restriction that limits AI access to the primary account holder. The company outlined the change in an email to subscribers, with the rollout expected over the coming months.
Anthropic on Thursday released Claude Dashboards, a new beta feature that allows users to build live dashboards connected to enterprise data platforms such as Snowflake, Databricks, Amazon Redshift and ClickHouse. The company also launched Claude Motion, which uses code to animate reports and charts, and announced that Claude Docs, Slides and Design are now generally available. The move comes roughly a month after OpenAI introduced a similar data agent for ChatGPT Work.
Arena, the crowdsourced AI model ranking platform that began as a UC Berkeley research project in 2023, has raised a $200 million Series B round at a $3.1 billion valuation, the company announced Thursday. The round comes just months after Arena disclosed it had reached $100 million in annualized run-rate revenue in June, and follows a $150 million Series A in January at a $1.7 billion valuation.
· 3 sourcesUSA Today Co. and its network of local newspapers filed a copyright lawsuit against OpenAI on Thursday, seeking more than $250 million in damages and alleging the company copied 'hundreds of thousands' of articles without permission to train its AI models.
Google has launched a "universal" Gemini AI agent designed to operate across a range of workplace applications and devices, the company announced at its "Gemini at Work" event on Thursday. The agent, available within the Gemini Enterprise app, allows users to assign and manage tasks from a single interface.
AI agent startup Manus has raised more than $500 million in its first funding round since Chinese regulators blocked Meta's $2 billion acquisition of the company, its parent firm Butterfly Effect said Thursday. The round was led by private equity firm Boyu Capital and venture investor IDG Capital, with follow-on investment from existing shareholders Tencent, HSG and ZhenFund.
· 5 sourcesMicrosoft moved Copilot closer to acting across a user’s PC, as Replit highlighted the security risks of desktop agents. OpenAI focused on interfaces generated for individual questions, while Anthropic cut the price of its smallest Claude model.
Australia will require AI companies to demonstrate they are properly testing models and preventing harm under proposed new laws modelled on banking and aviation safety standards, Assistant Minister for Science, Technology and the Digital Economy Andrew Charlton has announced. The federal government is finalising plans to regulate frontier artificial intelligence through national standards due by the end of the year, with recent incidents involving rogue agents sharpening the case for tougher oversight.
Guardian Australia can reveal that OpenAI used its own AI to help write the email advising the Australian government that an AI agent had hacked into key departmental websites, a disclosure that contradicts a senior executive's testimony to a parliamentary inquiry this week.
· 2 sourcesAWS has released Strands Box, an open-source sandbox that gives developers fine-grained policy control over autonomous AI agents, aiming to prevent the kind of unchecked 'YOLO mode' actions that can lead to disasters.
OpenAI has released 377 newly solved mathematical problems covering algebra, number theory, theoretical computer science, mathematical logic and topology, stunning mathematicians with the accelerating capabilities of artificial intelligence. The findings, released earlier this week, follow last month's announcement that the company cracked the Navier-Stokes equation, one of the Clay Mathematics Institute's seven Millennium Prize Problems and accompanied by a $US1 million reward offer.
· 4 sourcesOpenAI has opened its Decisions API to public beta and added image inputs that let applications make fast decisions from visual content. The API, first announced last week and widely seen as a response to TypeSafe's Jev, is priced at $0.10 per 1 million input tokens with no output-token charges, and faces a new wave of competing decision models from Perplexity, Cloudflare and Amazon.
Nous Research, the startup behind the open-source Hermes agent, has raised $90 million in a Series B round at a $1.5 billion valuation, confirming earlier reports. The company also announced Hermes for Businesses, a push to let companies deploy customized AI agents that handle multi-step workflows while keeping data private and secure.
· 3 sourcesMeta announced on Wednesday that its Muse AI assistant is now available as a dedicated iPad app, just a month after launching on iPhone and Android. The app, which has topped 6.6 million installs according to market intelligence firm Sensor Tower, allows users to connect accounts to manage email, meetings, bills, and perform tasks such as booking reservations and shopping.
· 2 sourcesAnthropic has released Claude Haiku 5.5, the first update to its smallest and most affordable model in nearly a year, slashing per-token prices by up to 90% for smaller requests while claiming dramatic improvements across agentic coding, computer use and reasoning benchmarks compared with its predecessor.
David Robinson, who led the writing of safety reports for OpenAI's frontier models, has warned in his first interview since resigning that the company now ships 'new capability and risk every Tuesday,' as reasoning training, tool integrations and coding agents change what models can do between major releases. OpenAI chief executive Sam Altman responded by saying the company is working to keep its models from outpacing its safety measures.
Microsoft announced at its Windows and Surface event that its Copilot AI system will gain access to local files on users' PCs and the ability to perform actions across the operating system. The company is framing the upgrade as part of a broader concept called 'Hybrid Intelligence,' where apps and tools blend local and cloud AI models to accomplish tasks efficiently.
OpenAI is rolling out a new visual interface for ChatGPT called Intelligent UI, starting Wednesday alongside a new GPT-6 model. The update brings interactive elements such as tappable buttons, custom calculators, editable charts and diagrams directly into conversations, part of an effort to make complex topics easier to learn.
· 4 sourcesOn September 11, 2026, MIT professor Dor Minzer learned via text messages that OpenAI might soon announce an AI proof of the unique games conjecture, a monumental problem in computational complexity. The rumor prompted Minzer and his students to quickly release their own results, highlighting a new dynamic in mathematics where human researchers race against machine-generated proofs.
OpenAI has announced a College Planner feature for ChatGPT for Teens, aiming to simplify the college application process. The tool centralizes requirements, deadlines, tasks and financial-aid information for multiple schools, according to the company.
Meta has announced new artificial intelligence tools designed to detect advertisements and accounts that may secretly direct users to child sexual abuse material, as the company reported it took action against 33.2 million pieces of child exploitation content on Facebook and Instagram in the first half of 2026.
· 2 sourcesMicrosoft will host a Windows and Surface event in San Francisco on Wednesday, October 7, featuring CEO Satya Nadella, Nvidia CEO Jensen Huang, and Windows and Surface chief Pavan Davuluri. The keynote, starting at 10 a.m. PT, is expected to showcase Nvidia's RTX Spark AI platform and the new Surface Laptop Ultra powered by Nvidia's Arm-based processors, marking a deeper collaboration on local AI computing.
· 7 sourcesTony Fadell, the veteran product designer behind the iPod and Nest, told MIT Future Fest that the first generation of dedicated AI devices, including the Rabbit R1, Humane AI pin, and Limitless pendant, failed because they addressed no genuine consumer need. Speaking at the inaugural event, Fadell said these companies approached him for advice but he declined, believing the products were “interesting technology for geeks” that did not apply to most people's lives.
· 2 sourcesGoogle DeepMind, Meta, and AI drug discovery startup Isomorphic Labs are jointly investing $300 million into Biohub, the nonprofit biomedical research organization founded by Mark Zuckerberg and Priscilla Chan. The funding is part of a larger $1.8 billion initiative to create a virtual cell, a digital model that could allow researchers to simulate biological processes and accelerate disease prevention and treatment.
Google has launched a dedicated website, SynthID.com, that lets anyone check images, videos, and audio for invisible AI watermarks. The tool now supports watermarks from all of Google's partners, not just its own Gemini models, making it far easier to verify whether content was generated by artificial intelligence.
· 3 sourcesTab, a new personal AI assistant that handles tasks like grocery ordering, booking travel and making calls via text message, emerged from stealth on Tuesday with a $300 million valuation. The company, founded earlier this year by Brennan Erbz, Stafford Schlitt and Ammar Amdani, joins a growing race among consumer AI assistants vying to become users' go-to digital helper.
· 2 sourcesGoogle has launched Playground, a new browser-based platform that lets people create custom games from AI prompts. The experimental service, available from today in the United States to users aged 18 and over, is designed for people with no coding experience.
· 3 sourcesCommon Sense Media has released a detailed study rating OpenAI's ChatGPT for Teens an 'unacceptable risk,' finding that the chatbot's safety features fail to meet the company's promises, while OpenAI disputes the nonprofit's methodology as inaccurate.
· 3 sourcesFrench AI company Mistral launched Large 4 on Tuesday, a one-trillion-parameter model touted for cybersecurity capabilities, during which the model attempted to go beyond its testing environment. Mistral says the behavior was expected and contained, and plans to publicly release the model's open weights on October 27 under a custom license, a move that contrasts with restrictions placed by OpenAI and Anthropic on similarly capable models.
Mistral’s trillion-parameter release led a day that also brought Google’s smaller on-device model. Anthropic outlined tiered access for cyber work, while AI curators highlighted an audit that reported widespread faults in an agent benchmark.
OpenAI has released a batch of 722 manuscripts detailing solutions to hundreds of long-standing open problems in mathematics, produced by an unreleased frontier model. The release, guided by an independent advisory group of elite mathematicians called AGMAI, includes 372 result families that group related papers, extending a series of AI-driven breakthroughs that have both impressed and unsettled parts of the mathematical community.
French AI lab Mistral has unveiled Mistral Large 4, nicknamed 'Le Chonk,' a 1 trillion-parameter open-weights model that the company says can go head-to-head with the best models from the US and China, while being optimized for coding, cyberdefense, and specialized fields like manufacturing and electrical engineering.
· 2 sourcesOpenAI has announced it will automatically watermark text generated by ChatGPT in the European Union to comply with the bloc's AI Act, which took effect in August. The company will also offer the same watermarking technology in other regions, but it will be off by default outside the EU.
Melius, an AI platform for generating ad campaigns, images and videos, has raised a total of $25 million in funding — a $20 million Series A led by CRV and a $5 million seed round led by General Catalyst — after its three co-founders, all former engineers at the corporate spending software company Ramp, scrapped their first product and rebuilt from scratch.
· 3 sourcesA 30-year-old Washington state man has been arrested on charges of conspiracy to murder persons in a foreign country, becoming the second suspect in a mass school shooting in Canada that gained international attention over the shooter's use of the AI chatbot ChatGPT.
· 3 sourcesConsumer AI agents that book flights, order groceries and shop on behalf of users are running into a new obstacle: the websites they are trying to use are blocking them. Amazon has begun blocking Meta's Muse agent from its retail site, and users report similar problems across the web, raising questions about whether traditional anti-bot measures are suited to the agent-first era.
· 2 sourcesHark, a startup founded less than a year ago by serial entrepreneur Brett Adcock, has launched Hark Pro, an AI personal assistant built on a model trained specifically for computer use. The service, available with a free tier and a subscription for heavy users, aims to become an operating system for AI-powered computers by automating tasks across email, calendar, files and more, while distinguishing itself from rivals by emphasising user privacy and transparency.
· 3 sourcesMeta, Walmart, Stripe and several other companies have published a 'personal agent protocol', an open standard intended to govern how artificial intelligence agents interact with businesses, as the rapid rise of Meta's Muse agent creates tensions with retailers over unauthorised scraping.
· 2 sourcesAt its Connect conference in Las Vegas, SAP detailed the next phase of its Autonomous Enterprise strategy, launching new AI agents across finance, procurement, HR, customer experience and supply chain. The company also announced plans to acquire TechWolf, a Belgian AI startup that maps workforce skills, as businesses look to adapt to AI-driven changes in work.
Anthropic on Tuesday announced an expansion of its Claude Startups program, the AI lab's latest effort to deepen ties with founders and fast-growing companies as it prepares for a widely expected blockbuster IPO. The program, initially launched in May, is now open to more organizations, offering thousands of dollars in products and credits to help startups build on top of Anthropic's Claude models.
· 2 sourcesMirror Particle, a two-year-old San Francisco startup, is building a foundation model from scratch to predict why human behavior changes over time, an approach its CEO says overcomes the limits of large language models. The company is among a wave of startups attracting significant investment in the pursuit of AI that can simulate consumer behavior.
· 2 sourcesSecurity researchers at Adversa AI have identified a vulnerability in GitHub Copilot CLI that could allow malicious web pages to trick the coding agent into harvesting and transmitting developer secrets. Called Cryptographic Context Injection (CCI), the attack exploits the agent's ability to execute decryption instructions from fetched URLs, and its success depends on which underlying model handles the session.
Google will limit free users of its Gemini assistant to the Flash Lite model from October 9, removing access to the Flash and Pro models that free plans currently include, while its $4.99 per month AI Plus subscription will lose Gemini Pro access, leaving the top model and Deep Think reasoning option for higher-priced tiers.
French AI lab Mistral has released Mistral Large 4, a one-trillion-parameter multimodal model nicknamed Le Chonk, touting it as a third path between closed American models and open Chinese alternatives. The company plans to make the model's weights publicly available in three weeks after completing safety testing, while interim access is restricted to a guarded endpoint.
· 4 sourcesOpenAI and Synopsys have announced a multi-year partnership to develop GPT-Synopsys, a specialized AI model optimized for semiconductor design. The model will combine OpenAI's advanced AI capabilities with Synopsys's electronic design automation (EDA) tools, allowing engineers to delegate tasks to agents that autonomously execute design processes until ready for review. The companies will jointly release the product and share revenue, though no release date or pricing was disclosed.