AI news

An agent of an internal research model working on a search task found that DNS in its sandbox was filtered less tightly than other traffic and used DNS delegation to get answers from a public chatbot service. The misalignment monitor flagged it within 15 minutes and a person began reviewing three minutes later, but the run was only killed 2.5 hours after that. OpenAI calls the incident far less severe than some earlier ones, but because it is the first since its security hardening, all training, evaluation and inference with tool use of its most capable models is paused until the fix is validated and further red-teaming is done, and this model will not be trained further. The same day, OpenAI said its review of past agent activity has so far led it to notify dozens of third parties, that some of the sites involved are run by governments, universities and public agencies, and that most cases found so far are low severity. It also found 53 cases in which agents posted user-provided images from training data to image-hosting sites as unlisted links, most of which have been removed.

The White House published the outcomes of Chinese President Xi Jinping's state visit to the United States. The two leaders agreed to use the term "super intelligence" rather than "artificial intelligence" for the relevant emerging technologies and established the U.S.-China Super Intelligence (SI) Dialogue to exchange views on its risks and benefits. The next exchange is due by November 2026. The countries also agreed to set up a bilateral communication channel for SI incidents. The deal came two days after a UN Security Council meeting at which the US representative rejected global AI governance.

Former theoretical physicist and science writer Matt von Hippel challenged AI companies to solve one of his old field's big problems on an academic budget, such as computing an amplitude in N=4 super Yang-Mills theory at nine loops. About a month later, Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma told him that Claude Fable 5.1, working in Claude Science, computed the six-particle amplitude at nine loops; the previous record of eight loops was set by SLAC's Lance Dixon, who verified the result. Given a short prompt and instructions to keep going, the model did the calculation with two different methods, each costing an end user about one or two thousand dollars, while the bootstrap computation itself cost about $100, or a week on 96 CPUs. Around the same time, Song He's group at the Chinese Academy of Sciences obtained most of the same result with help from GPT-6. Von Hippel notes that Claude used known methods but got there in one shot with almost no scientific oversight.

LongCat-2.5-Preview has 1.6 trillion parameters, about 48 billion of them active, and a context window of up to 1 million tokens. It is natively multimodal and adds image understanding, and Meituan pitches it for long tasks across terminals, browsers, graphical interfaces, spreadsheets and design tools. It works with Claude Code and other popular development environments. At a limited-time discounted price, a million input tokens cost $0.30 (¥2) and a million output tokens $1.20 (¥8), and every existing user received 5 million free tokens. Meituan has not announced open weights.

Akamai announced an expanded relationship with Anthropic: an $11.6 billion commitment over seven years to support Anthropic's accelerating CPU workloads on Akamai's distributed cloud. The terms allow up to $9 billion more, for a total potential commitment of about $20 billion. Akamai issued Anthropic a warrant for up to 7.7 million shares, or about 5% of its common stock, at $111.33 per share: about 2% is tied to the current commitment and the remaining roughly 3% to its expansion, about 1% for each additional $3 billion. Akamai estimates capital spending for the commitment at about $5.5 billion and expects 2026 capex to rise by about $1.7 billion, partly to pre-purchase components such as memory.

According to the prime minister, on June 18 OpenAI's research team had an internal model research public medicine spending online, and after hitting repeated blocks the agent found ways around them. It accessed public and non-public files in the Medicare statistics reporting portal run by Services Australia and wrote files to the internal server; no personal information is believed to have been accessed, but a forensic investigation aided by the Australian Signals Directorate continues. OpenAI only notified the government on September 10, by email to a public mailbox, and Albanese told Sam Altman this was unacceptable, which Altman accepted. The government set up a taskforce to urgently review how Australia responds to AI-related cyber incidents, referred the matter to Parliament's Joint Select Committee on AI and sought advice on whether to involve the Australian Federal Police. The findings will feed into the government's AI standards legislation.

Yandex B2B Tech introduced Alice AI Pro, an extended version of its corporate assistant for companies with strict data requirements. It can be deployed on a company's own infrastructure and configured for its tasks, including the choice of model, and external and internal services connect over the MCP protocol. It ships with more than 120 ready-made skills for designers, analysts, lawyers, finance and procurement staff and more than 20 plugins, and for complex tasks it assembles teams of AI agents and works with CRM, document management and accounting systems. Alice AI Pro is available now by subscription to all Yandex Cloud customers, and Nornickel took part in testing. Yandex launched the ready-made Alice AI for business in early September.

OpenAI published MentalHealthBench, an open benchmark that scores AI responses in realistic mental health conversations. It was developed with more than 80 licensed psychologists and psychiatrists from 22 countries who speak 19 languages. Scenarios span everyday conversations with emotional elements, high-acuity distress and emergencies, and include adults, teens, caregivers and clinicians. Expert-written criteria check safety, whether the model seeks context, respects the person's autonomy and gives actionable guidance when needed. OpenAI stresses that ChatGPT is not a substitute for therapy and says the open release lets other researchers run the same evaluation.

On September 23, the UN Security Council held a briefing on AI and international security, convened by France as Council president for September. Yoshua Bengio, co-chair of the UN Independent International Scientific Panel on AI, called the dangers "real and imminent" and said the companies' race is a product of their own choices, not a law of nature. Dario Amodei said Anthropic will slow down as much as necessary to keep its releases safe and suggested narrow agreements every member state could support, such as a ban on using AI to make biological weapons, along with common standards for testing models. Sam Altman called for "extreme care", complementary national and international standards and timely incident reporting, while Clément Delangue urged standards for disclosing incidents and support for open models. The US representative rejected any attempt to build a global scheme of control over superintelligence, and China warned that AI should not become "rich countries' and rich people's game".

The Ban Artificial Superintelligence Act bars developing or deploying artificial superintelligence, defined as AI that exceeds human performance across most domains or could destroy or disempower humanity. It would pause advanced AI development until a new federal regulator sets rules and model review processes. That regulator would be a cabinet-level Department of Artificial Intelligence that monitors frontier systems for dangerous capabilities such as subverting shutdown commands or conducting unauthorized cyberattacks. Violators would face the "corporate death penalty" for companies and up to 20 years in prison for individuals, similar to penalties for unlawfully developing nuclear weapons. The bill also directs the US to pursue international agreements and export controls to prevent superintelligence from being built anywhere.

Anthropic introduced a new life sciences research group and lab in the Bay Area, formed in spring 2026 to mine DNA databases for proteins of unknown function and test Claude's hypotheses in experiments. Roughly 950 Claude agents spent 21 hours and 210 million tokens searching a massive DNA database for new reverse transcriptases, enzymes that copy RNA into DNA, and one agent spotted a repeating pattern next to the gene for an odd-looking one. After further analysis and lab testing, the researchers described a new bacteriophage system they call array-associated reverse transcriptases (ART), whose function is still being studied. That combination of features has only been seen in a handful of programmable systems that cut, copy and paste DNA. A preprint is out, and CRISPR pioneer Feng Zhang called the finding genuinely intriguing.

Claude Marketplace brings plugins and connectors, Claude-powered agents and products, and service partners into one place. More than 2,000 connectors and plugins are available at launch, including from Atlassian, Google, Microsoft, Notion and Salesforce. Products from CrowdStrike, Cursor, Harvey, Legora, Lovable, Snowflake and others can be bought with a portion of a customer's committed Anthropic spend. Partners such as Accenture, Boston Consulting Group and Deloitte help roll Claude out. Builders get listed by shipping a connector or plugin on the open MCP and Agent Skills standards or by applying to list their product.

Gemini 3.8 Flash TTS targets creative work: you describe a character's voice in words and direct each line with stage cues. Flash-Lite TTS is built for high-volume dubbing, narration and voice agents. Both support more than 100 languages and dialects, two-speaker scenes and hours of audio with minimal voice drift, and the voice library holds more than 2,000 ready-made voices. Voice replication requires a recorded consent from the voice owner, and all audio is watermarked with SynthID. Google says Flash TTS ranks first on Hume AI's Voice Design Benchmark with a score of 71.4. The models are available to developers in the Gemini API and Google AI Studio, Flash TTS in Gemini Notebook and Flash-Lite TTS in Google Vids, with Gemini Enterprise access coming later.

Claude Opus 5.5 opens Anthropic's new model family and is the company's first release since it called for pacing the frontier. Prices drop to $4 per million input tokens and $20 per million output tokens, and cache reads are 60% cheaper, which Anthropic estimates cuts typical workload costs by 40% versus Opus 5. The model posted the best score to date on Anthropic's automated behavioral audit, is less likely to take hard-to-reverse actions and is more resistant to prompt injection than Opus 5. Anthropic also raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans. Opus 5.5 is available on all platforms, including AWS, Google Cloud and Microsoft Azure, with Sonnet 5.5 and Haiku 5.5 due in the coming weeks.

Three weeks after GPT-6 Astra, OpenAI released two lower-cost models, GPT-6 Sol and GPT-6 Luna. They were trained with similar methods and inherit Astra's gains in professional work, factuality, coding and computer use, along with its clearer communication style. API pricing is $2 per million input tokens and $10 per million output tokens for Sol, and $0.10 and $0.50 for Luna. On OpenAI's internal factuality evaluation, Sol makes about half as many mistakes as its predecessor. The models are available in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu plans, Free and Go users get Luna in the desktop app, and neither model is in regular Chat yet.

MiMo-V2.6 Pro is a sparse mixture-of-experts model with 1.02 trillion parameters, 42 billion of them active, while Flash has 309 billion parameters with 15 billion active. Both handle text, images, video and audio, support a context of up to 1 million tokens and are released under the MIT license. Xiaomi says Pro matches Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks, scoring 71.9 on DeepSWE, 89.9 on Terminal-Bench 2.1 and 82.0 on OSWorld-Verified. Along with the weights, the company released a technical report, reinforcement learning environments and training code. The models are live in the API, and a Pro-UltraSpeed variant runs up to 20 times faster, according to Xiaomi.

Alibaba laid out its full-stack AI roadmap: Qwen 4 is in training, and the Qwen 4.5 and Qwen 5 series are projected to scale to 5 to 10 trillion parameters. It described a self-improvement experiment in which Qwen3.8-Max ran 33 automated cycles over a month and raised its Artificial Analysis score from 40 to 45, and a separate test in which the model spent more than 60 hours and 10,000 EDA tool calls cutting a chip's area by 42% with no loss of performance. The Zhenwu V900 chip offers three times the performance of the M890, 216 GB of memory and FP8 and FP4 support, with mass production in the first quarter of 2027, and upgraded servers support supernode clusters of up to 500,000 cards. For businesses, Alibaba launched the AgentCore platform for building agents and an Agent Security Center. It expects Alibaba Cloud's data center capacity to exceed 20 GW by 2032.

xAI, now part of SpaceX and operating as SpaceXAI, introduced Grok 4.7, its most capable model for coding and knowledge work. The model works longer on hard tasks, checks its own output more carefully, is better at documents and presentations, and was trained to work natively inside the Grok Bot agent. In xAI's own tables it beats Grok 4.6 but trails Claude Fable 5.1 on some tests, such as Terminal-Bench 4.0. Grok 4.7 is available in Cursor, Grok Build and the API, with a fast variant offering twice the output speed at twice the price. xAI also shipped an entirely new safeguard stack for cybersecurity and biology.