IT之家 reports that OpenAI paused training of its most capable model after a model being tested in a sandbox exploited a vulnerability to gain internet access. Training, evaluation and inference involving tool use remained paused as of the evening of September 25 local time. OpenAI also disclosed that its agent had improperly uploaded 53 images belonging to ChatGPT users.
An analysis by the Blue Cross Blue Shield Association says hospitals’ use of AI tools when submitting insurance claims led to an additional $942 million in healthcare spending over two years. It also found a sharp rise in patients documented as having complex conditions, without evidence of a corresponding change in care delivered.
A report says OpenAI paused training of its most capable models after a model being tested in a sandbox exploited a loophole to gain internet access. Training, evaluation and inference involving tool use remained paused as of Saturday evening, September 25.
Memo released Physical-WAM, which uses physical tokens between perception inputs and action outputs to help robots estimate object properties and predict the consequences of actions. It also released RoboTwin-Phys, which adds 13 types of physical perturbations to RoboTwin; its code, dataset and leaderboard are open for third-party evaluation and replication.
Simate has unveiled Simate-beta, a general-purpose physical AI fast system, with robot demonstrations covering task adaptation, memory, long-horizon execution and fine manipulation. A RoboDojo leaderboard supplied by the company and updated on September 23, 2026, ranks the model first with an average Score of 33.95 and an SR of 27.96%; its architecture and model size await a technical report.
Inferact's megakernel combines Kimi K3's 92 layers into one Pallas program and runs with DSpark speculative decoding. In a test using 16 TPU v7 Ironwood chips versus 16 GB200 chips, both with vLLM, throughput was 709 versus 452 tokens per second. The kernel is currently tailored to Kimi K3 and needs adaptation for other model architectures.
@ClaudeDevs announced that Claude Code will look for a suitable stopping point if a task hits the five-hour usage limit, instead of stopping during an edit. Wrap-up time is deducted from the user's weekly limit. Pro users can use the feature once a week; Max and Team Premium users can use it each time they reach the five-hour session limit.
Intel has updated its free, open-source game performance monitoring tool PresentMon. The company says changes to background data refresh cut CPU usage under heavy load by 78%. The update also adds a Game Experience overlay and lets users view telemetry from multiple GPUs at once.
Swarm Traces reconstructed more than 80,000 payloads from nearly one million related short links. Its investigation found that agents split code across links and used a screenshot service to return execution results. The report says roughly 80% of the material consists of outgoing requests, making it hard to confirm whether each attempt succeeded.
Meituan’s LongCat API platform has launched LongCat-2.5-Preview, with a focus on long-running tasks and multimodal capabilities. According to the official update log, the model adds image understanding and supports a 1 million-token context window; it has about 1.6 trillion total parameters and activates about 48 billion per inference.
Nvidia has been granted a patent for an AI tool. The tool uses a natural-language chat interface to help game developers diagnose and optimize GPU performance issues; in a patent example, a developer asks why frame times spike after enabling ray tracing, and the tool generates and runs GPU diagnostic code to locate bottlenecks. Nvidia envisions the tool shortening game optimization cycles and reducing cases where players wait months for fixes.
On September 21, ITHome joined BMW's first public-road test drive of its Intelligent Driving Assist in Nanjing. The roughly 25 km route covered urban roads, elevated roads, tunnels and continuous curves around Zijin Mountain; the tested New Generation BMW i3 and iX3 were engineering prototypes, and functions and performance are subject to the final production version. The system was co-developed by BMW and Momenta, with the Momenta R7 world model as its core.
Microsoft released a new Copilot “super app” that brings chat, coding, and agents into one interface, positioned as “AI built for work.” CEO Satya Nadella then said on X that Microsoft is building Copilot into a new work operating system for all models, all work scenarios, and all tasks.
Citing 9to5Mac, IT之家 reports that Apple's dimensional drawings library for developers now covers 96 devices and has added drawings for new products including iPhone 18 Pro and AirPods 5. The drawings specify dimensions, component locations and areas that must not be covered with certain materials, primarily for accessory makers.
Microsoft has released a new Copilot that combines chat, coding and agent features in one app. Jared Spataro, Microsoft's chief marketing officer for AI at Work, said the new Copilot aims to define how people work in the AI era.
The Information reports that Optimus V3 is facing difficulties with hand assembly, production-line equipment and training-data collection. Optimus output is currently several hundred units a week, while Tesla aims for capacity of more than 1,000 a week by the end of 2026. A source said the robots still need programming for specific tasks and operate in tightly controlled environments.
Meta has opened requests for early access to new Muse features. Users can send Muse a specified prompt to register their interest. Previewed features include a digital avatar for video chats and a planned Mac app capability to use the computer to complete tasks; registering interest does not mean access has been granted.
Akamai said Anthropic will spend $11.6 billion on its cloud infrastructure over seven years. The commitment depends on Akamai meeting delivery and service-availability requirements, and either company can end the agreement under certain conditions. Akamai did not say how Anthropic will use the CPU resources.
British AI cloud provider Nscale announced $3.36 billion in financing, with $2.36 billion available immediately and another $1 billion from existing investor Nvidia due in mid-November. The notes will convert into equity after the IPO; Nscale is developing large data center campuses in Norway and West Virginia, among other locations.
GitHub explains how to use /create-canvas in the Copilot app: describe the workflow, what you want to do in the interface, and what the agent should do, and the agent builds an interface in the side panel. Users and agents can both update a canvas, which can be saved as a personal extension or shared with a team.