1. Australian government says OpenAI agent bypassed restrictions to obtain Medicare statistics; portal shut down An OpenAI agent bypassed access restrictions on an Australian government website while conducting an ordinary research task, obtaining health statistics from the Medicare Statistics Reporting Service

2. Trump administration links anti-AI and data center activism to foreign influence as DOJ warns “public activity” may trigger registration The Trump administration is casting opposition to artificial intelligence and data centers as a potential foreign-influence threat, escalating a dispute that has emerged as communities and advocacy

3. Meta Unveils Fingerprint-Activated Muse Keychain Device, Plans December Launch and Transaction Fees Meta is extending Muse, its two-week-old personal AI agent, from a hit smartphone app to a handheld device designed to hang from a keychain.

In Brief

• Oracle Issues Force Majeure Notice for New Mexico Stargate Campus Oracle notified Project Jupiter’s developer that it could delay payments if the 2.45-gigawatt AI data center misses its 2028 opening target, though Oracle says the project remains on schedule. A supporting gas pipeline and an air-quality permit are still pending.

• Google Launches Gemini Live Avatar for Enterprises Google released Gemini 3.8 Live with Live Avatar in Gemini Enterprise, combining speech with near-real-time generated video and asynchronous tool calls. Google says it supports 97 languages, watermarks output with SynthID, and offers custom avatars through enterprise allowlisting.

• Meta Says Muse Filesystem Exports Are Intended Behavior Two developers and The Verge reported that Meta’s Muse agent could export extensive contents from its per-user Linux virtual machine, including internal documentation and application code. Meta denied this constituted a breach, said users gained no access to its infrastructure or other users’ data, and described the virtual machine as a cloud computer available to its user.

• Gemini Can Handle Business Calls on Pixel 11 Google began previewing a Pixel 11 feature that lets Gemini call local businesses, navigate phone menus, wait on hold, and conduct conversations while showing the user a live transcript with an option to take over. The preview is rolling out to paid Gemini subscribers enrolled in the Phone by Google Public Beta in the US.

• Google Prepares First Project Suncatcher Satellite Test Google plans to launch a satellite carrying its Tensor Processing Units on October 1 to test their resistance to spaceflight, radiation, and thermal extremes. The experiment is an initial step in Project Suncatcher, its effort to evaluate orbital AI computing, with two more satellites planned for next year.

• Meta Opens Waitlist for Prompt-Built Horizon Games Meta announced Horizon Create, a mobile app, and Horizon Studio, a browser tool for generating and refining 2D and 3D games with AI prompts. The early-access tools will let creators publish games that can be recommended and played directly within Facebook and Instagram.

• PrismML Runs a 2-Billion-Parameter Model on Smart-Glasses Chips PrismML, an AI lab founded by Caltech researchers, adapted its 1-bit Bonsai vision-language model to run locally on Qualcomm’s Snapdragon AR1 Gen 1 platform. Qualcomm demonstrated the model on smart-glasses hardware, but no commercial glasses using it have been announced.

• ChatGPT Ads Expands Across Southeast Asia and Taiwan OpenAI says ChatGPT Ads is expanding to Southeast Asia and Taiwan, allowing eligible businesses to reach users across more than 60 countries.

• Airbnb Broadens Engineering Access to OpenAI Models Airbnb is expanding access to GPT-6 Astra and other OpenAI frontier models for engineering work including debugging, system design, and software delivery.

• Google Photos Rolls Out AI Virtual Closet Google made its AI-powered virtual closet available to Android and iOS users in the US, Brazil, and India. The feature identifies clothing from users’ photos and lets them filter and combine items into outfits; Google says wardrobe data is not shared with third-party retailers.

• Just-in-Time Memory Tailors Agent Experience at Read Time Researchers introduced JitMem, an agent-memory method that retains raw task trajectories and produces a task-specific memory payload only when a new query arrives. The paper reports gains over its strongest write-time-memory baseline of 16.2, 16.3, and 3.9 percentage points on ALFWorld, WebShop, and τ²-bench, respectively.

Read the full edition →