Today's Flow

On September 25, 2026, at the forefront of the artificial intelligence industry, a clear trend is emerging: efforts to translate model intelligence into real-world interaction and execution capabilities. At the center of this movement lies agent functionality that communicates directly with external systems on behalf of users. Google holds the Pixel 11 device and utilizes Gemini subscriptions among U.S. A test has begun in which users have Gemini directly calling companies to carry out business tasks. This marks a practical turning point, signaling the entry of generative AI—previously confined to text interfaces—into real-world business touchpoints via actual communication networks.

Alongside the expanded execution capabilities of these agents, governance tools are being strengthened in development environments to safely control autonomous AI behavior. GitHub has introduced local sandboxing for its Copilot applications to proactively block the risk of unexpected commands accessing files, networks, and credentials, while also enhancing code review Expanded request and configuration options to enhance control in enterprise environments.

Alongside this, practical applications are advancing rapidly in both everyday services and research and infrastructure domains. Google Photos has fully expanded the deployment of its virtual closet feature—which identifies clothing from users’ photo libraries to create a curated wardrobe—from Android to iOS. In terms of model lightweighting and acceleration, Liquid AI accelerates vision-language model processing. It introduced a new approach, and NVIDIA unveiled tools that significantly shorten robot simulation and learning workflows. OpenAI is also continuing to expand community-driven technical capacity support as it celebrates the second anniversary of its Academy.

Key News

Google Gemini Enterprise Phone Agent: Pilot Testing Begins for U.S. Pixel 11 Paid Subscribers

Google has begun officially testing an AI-powered calling feature that makes phone calls to businesses on behalf of users. According to a TechCrunch report, this feature is initially being offered to Pixel 11 device owners who are paid subscribers of Gemini in the United States. Instead of users having to call businesses themselves, Gemini places the calls and handles the necessary communication. It is implemented as a proxy service. Although the currently released official materials do not specify in detail the scope of supported companies or the detailed scenarios for currency handling, it is interpreted as a strategy to verify initial stability by integrating the currency feature with the latest flagship devices and paid subscription plans. The smartphone's built-in assistant executes actual voice-based calls. It is considered an important step in evolving into an agent.

Google Photos Virtual Closet: Expands to iOS Platform Following Android

The virtual closet feature in Google Photos, inspired by the movie 'Clueless,' has officially expanded to the iOS platform following its Android launch. This service uses artificial intelligence to analyze everyday photos stored in users' photo libraries and automatically builds a virtual wardrobe. It was first rolled out to Android device users last June. Following its phased rollout, this update has expanded the service to cover both major mobile operating systems. By integrating personalized fashion and lifestyle management features into photo management tools, it represents a significant advancement in leveraging artificial intelligence for everyday visual data.

GitHub Copilot Local Sandboxing: Mitigating Security Risks Through System Resource Isolation

GitHub has officially introduced local sandboxing for its Copilot application. This measure aims to mitigate unintended system changes or potential harm that could arise when AI agents in the development environment execute commands on the user's local machine. Within the local sandboxing environment, the internal file system of the machine is... Access to network resources and stored credentials is strictly restricted. The Copilot app supports users in configuring these settings directly, with a focus on institutionally mitigating the security threats that agent-based automation tools may pose to developers' local development environments.

Advanced GitHub Copilot Code Review Configuration: Support for Personalization and Enterprise-Wide Default Control

GitHub Copilot's code review feature has been significantly revamped and officially released to the public. This update expands the supported Copilot plans, introduces enterprise-level default policy settings, and adds dedicated customization options tailored for individual developers. Through these enhancements, teams and organizations can now streamline their code review processes by It is now possible to standardize the rules and feedback mechanisms that artificial intelligence should follow, allowing developers to finely tune their review request processes to match their own working styles. This lays the groundwork for ensuring the reliability and consistency of AI-driven code reviews across collaborative workflows.

Liquid AI's Vision-Language Model Acceleration: Release of LFM2.5-VL-DSpark

Liquid AI has announced LFM2.5-VL-DSpark via the Hugging Face blog, a technology designed to accelerate vision-language model processing. It focuses on resolving computational latency and bottlenecks that occur when large-scale multimodal models simultaneously process visual information and text. Currently, the official materials released do not include detailed architectural specifications or Although specific benchmark metrics were not explicitly stated, it laid the groundwork for efficiently integrating multimodal AI into various applications requiring real-time responses by dramatically accelerating visual-language integrated reasoning.

NVIDIA Physics Simulation Acceleration: Innovating Robotics Learning with Warp and MjWarp

NVIDIA introduced methods for accelerating robot simulation and machine learning workflows by leveraging the high-performance physics framework Warp and MjWarp, via its Hugging Face technical blog. Robot systems that must operate in physical environments need to learn control policies through large-scale iterative simulations in virtual environments, a process during which We presented a pipeline that maximizes simulation efficiency by applying a GPU-based accelerated computing framework, providing a technical blueprint for robotics researchers and engineers to shorten the development cycle from simulation to actual hardware deployment.

OpenAI Academy’s Second Anniversary: Expanding Global Community Support for AI Literacy

OpenAI Academy has reached its second anniversary since its launch. Through a commemorative announcement, OpenAI reflected on its operational achievements over the past two years and emphasized its goal of delivering AI practical skills to more global communities and diverse groups. By bridging the technology gap, it aims to empower each community to independently adopt AI tools to solve problems. It is committed to continuously expanding its educational infrastructure and support programs. This move reaffirms the resolve not only to develop technology but also to cultivate a broad base of human resources capable of embedding these advancements in practical applications.

Key Points to Watch

Legal and Ethical Governance of Autonomous Agents Acting as Phone Representatives: Disclosure of Call Parties and Verification of Interactions

As Google's testing of its Gemini enterprise call-handling capabilities intensifies, the key focal points are how identity disclosure obligations, call recording requirements, and privacy regulations will be concretely defined for artificial intelligence systems when they engage in voice conversations with human recipients. In particular, attention is centered on determining liability in the event of information transmission errors or decision-making confusion that may arise during calls. Industry standards are required. It is necessary to closely monitor the real-world feedback accumulated during the initial Pixel 11 and paid subscriber pilot programs, as well as Google's subsequent guideline announcements.

Isolation of Developer Agent Privileges and Enterprise Security Standards: The Ubiquity of Sandboxing

GitHub Copilot's local sandboxing and granular code review controls are becoming essential safety nets for the proliferation of autonomous coding agents. Key future developments will focus on how well these isolated environments can fully meet corporate security compliance without hindering developer productivity, across various local development setups. And during the process of linking with corporate policies, we should monitor the adoption speed of enterprise customers and the expansion of additional permission control features.

Convergence of Multimodal Lightweighting and Robotics Physics Simulation: Advancing Toward Embodied Artificial Intelligence

Liquid AI's vision-language model acceleration and NVIDIA's robotics simulation acceleration tools are laying the technological foundation for the era of "embodied artificial intelligence," where AI understands and acts in the physical world. When real-time reduction of visual information processing latency is combined with efficient large-scale physics simulation, it enables... It remains to be seen how much performance improvement can be achieved, and the subsequent application cases in the open-source community and industry are drawing attention.

Source

• TechCrunch AI: https://techcrunch.com/2026/09/24/google-tests-letting-gemini-make-phone-calls-initially-for-us-pixel-owners/
• TechCrunch AI: https://techcrunch.com/2026/09/24/google-photos-clueless-inspired-virtual-closet-is-now-available-on-android-and-ios/
• GitHub Changelog: https://github.blog/changelog/2026-09-23-local-sandboxing-in-the-github-copilot-app
• Hugging Face: https://huggingface.co/blog/LiquidAI/lfm2-5-vl-dspark
• GitHub Changelog: https://github.blog/changelog/2026-09-23-copilot-code-review-more-ways-to-request-and-configure-reviews
• Hugging Face: https://huggingface.co/blog/nvidia/how-to-use-nvidia-warp-and-mjwarp
• OpenAI: https://openai.com/index/two-years-of-openai-academy