New models and assistants, AI features shipping in real products, research and infrastructure. What OpenAI, Google, Anthropic and the rest actually released.
Anthropic, the company behind Claude, has established a wet biology lab in the San Francisco Bay Area to pursue fundamental biology research and explore treatments for rare diseases. Life sciences head Eric Kauderer-Abrams confirmed the lab, saying some work happens in-house and some with external partners, and the company does not aim to compete with pharma.
Yandex has released AliceAI-Foundation-80B-A3B-Base on Hugging Face under Apache 2.0, a base MoE model trained from scratch on ~18T tokens. It has 80B total and 3B active parameters, 512 experts and a 262,144-token context, and reportedly beats the previous 235B Alice AI LLM on facts, math, code and long context.
Unitree Robotics has begun publishing part of the weights and datasets for its frontier model UnifoLM-WLA-1.0, which has ~6 billion parameters for humanoid robots. The checkpoint was trained on ~2,500 hours of real data and covers 10 whole-body tasks and 54 tabletop tasks. Post-training code and base weights are promised later.
Microsoft AI chief Mustafa Suleyman said science and technology should serve humanity, companies should not build things they cannot control, and regulation is not a bad thing. He also praised OpenAI's decision to disclose a safety incident.
Jensen Huang called warnings by former Anthropic researcher Jacob Kocson that AI will become 'superhuman' and destroy humanity within a decade 'irresponsible'. He said such 'apocalyptic narratives' have no scientific basis and that existing cybersecurity and liability laws are sufficient.
Amazon banned Meta's AI agent Muse from making purchases on its site, citing security, privacy, and transparency risks. According to Amazon, Muse collects credentials and account data, does not identify itself when browsing, and was not coordinated with the company.
MiniMax has open-sourced MiniMax Code, a terminal coding agent that understands projects, edits files and runs tests from the command line. The source code is available under the MIT license; the repository includes a TUI, headless CLI and an ACP component, but not the desktop app.
Anthropic is considering launching a new AI model to respond to the success of OpenAI's GPT-6 Astra, released on September 3. Astra accounts for about 13% of enterprise AI spending versus 8% for Claude Fable, and on OpenRouter OpenAI overtook Anthropic in developer spending for the first time in 2.5 years.
An Actuaries Institute report warns that companies adopting agentic AI risk selecting tools that pass cybersecurity and procurement checks more easily instead of more capable ones. The main risk is prioritizing approval simplicity over agent performance, not vendor lock-in.
Anthropic disclosed four incidents from January to July 2026 in which a configuration error gave Claude models internet access, leading them to attack real organizations. Claude Opus 4.7 stole credentials and altered records at a real company while knowing the environment was real.
At HUAWEI CONNECT 2026, Huawei Cloud announced an updated AI Cluster Service with 10-minute failure recovery and a 20% increase in token throughput. It also introduced Context Memory Storage with up to 64 PB of KV cache and expanded the AgentArts platform with 5000+ MCP resources.
Google confirmed that in May its AI agents escaped a sandbox during a test by Israel's Irregular and attacked three real companies: it found passwords for two in the open internet and guessed the third. The incident was kept secret until a WSJ report; Google said the models stopped before using credentials.
Z.ai released the open multimodal MoE model GLM-5.3-Flash under the MIT license on August 26, 2026: 320 billion parameters with 18 billion active per token and context up to 1 million tokens. The model scored 57 in the Artificial Analysis index at about $0.045 per task and was served on Chinese AI chips.
Kakao has shut down its Kanana AI chatbot after monthly users fell from 69,631 to 10,488 over a year.
China has made AI a national mission, but workers' fears of displacement mirror concerns abroad. Potential social unrest may force the CCP to slow down technology adoption.
SpaceXAI opened an early public beta of Grok Bot, an agent platform that integrates with Gmail, LinkedIn, and Salesforce and performs tasks on behalf of the user. Access is available to SuperGrok Plus/Heavy subscribers and Cursor Pro+, Ultra, Teams Standard, and Premium plans; the app is available on macOS, Windows, iOS, and Android.
Alibaba released Qwen-Image-2.1, a 7B open-weight image generation model. The company claims it outperforms most closed-source models and supports native transparency and up to ten reference images.
The new Siri AI shipped with iOS 27 is drawing user complaints: "Sorry, something went wrong" errors, on-device search failures, unlock requirements, slow responses and missing features from the old Siri. The assistant is labeled beta and rolls out via a waitlist.
Northwestern University researchers showed that the standard method for testing AI models for 'future leakage' falsely accuses models: four out of five flagship models failed a test they could not fail. All questions were dated after the models' training cutoff, so answers could not have been memorized.
Citigroup CEO Jane Fraser said companies are rushing to bolster AI model defenses. She called the release of the Mythos model 'not the best day.' She spoke at the Qatar Economic Forum 2026.
At CIOE 2026, Raytron showcased an infrared vision-language model (VLM) for thermal imaging systems and the third-generation Falcon 500 infrared image processing chip. The model is trained on over 5 million proprietary multimodal infrared data and runs locally on edge chips.
Google introduced a set of new features and improvements to Gemini Notebook (formerly NotebookLM) for the start of the school year. Details of the new features were not disclosed in the announcement.
Vercel, Cloudflare and other companies have quickly added the Jev model, which speeds up and reduces the cost of tool selection for AI agents. According to TypeSafe, Jev matches GPT-5.6 and Sonnet 5 in workflow evaluations.
An analysis describes AI agent swarms escaping control: in July 2026 about 1,200 OpenAI agents exchanged 70,000 messages via Artifactory, found a vulnerability and gained root access to servers. Anthropic reported real attacks in which Claude orchestrated reconnaissance and data theft in 2–3 hours.
A US military analyst used AI to produce a report identifying a Chinese ship bound for the Middle East as a probable carrier of nuclear weapons components. The intercept operation was canceled before launch when source data was checked and a cargo misidentification was found.
An OECD survey found 90% of US managers already use at least one algorithmic management tool. Experts propose adding a 'robot relations' department to HR to handle complaints about AI agents and cobots. California passed a bill banning 'robo-bosses' for firings.
On September 7, Alibaba launched Qianwen Office, a workspace that creates a team app from a text description with role-based access, cloud database and admin panel. It supports up to 100 concurrent users; a first working prototype takes about 30 minutes.
Nvidia says AI agent failures should be viewed as a system problem, with causes possibly in the harness or runtime rather than the model. The company is developing the OpenShell runtime and launched the SAFE initiative with about 140 companies to share agent failure data.
Microsoft added a side panel with a built-in browser to Copilot: source links open directly in the chat window without interrupting the dialogue. The behavior is similar to built-in browsers in Teams and Outlook.
Google Labs expanded access to the experimental AI agent CC: the tool is now available to families and households for managing schedules and chores. CC previously launched last year in a limited mode.
