Anthropic posted a 2026 Usage Policy last Thursday that takes effect November 12, with a ban on using Claude as infrastructure for fake accounts, fabricated sites, and influence campaigns. NVIDIA showed developers directing GPT-6 Astra and Claude Fable 5 in natural language while Omniverse libraries supply the physics, rendering, and sensors. Google Cloud's Java MCP Toolbox client reached a stable 1.0 for AlloyDB and Spring shops. OpenAI is adding a College Planner to ChatGPT for Teens, plus a student advisory council at Boston Children's Hospital.

In the news

Anthropic 2026 Usage Policy

Anthropic published a new Usage Policy on October 8 that takes effect November 12. Most of the rewrite is meant to make existing rules explicit as Claude takes on longer, more independent work.

The section people will actually feel is a new consolidated rule, Do Not Engage in Deceptive Campaigns or Artificial Activity. It covers fake accounts, fabricated sites, and the tools and infrastructure for running influence campaigns, political or commercial. The elections section is now titled Do Not Undermine Democratic Processes. Anthropic dropped a blanket ban on personalized vote and campaign targeting so civic groups and election officials can still use Claude for things like ballot notices, while deception and misuse of voter data stay prohibited.

Weapons rules now explicitly cover guidance and control software and arming drones or other autonomous vehicles. Surveillance language bans non-consensual tracking, recommending who to arrest, and building surveillance tools. High-risk health, legal, and finance work still needs a qualified human in the loop plus disclosure that AI was used.

A new hardware clause, after the Model Hardware Standard, requires an operator who can stop equipment and a safe state if Claude disconnects. Anthropic also added a ban on sustained, needless cruelty toward its models. That ban is meant for extreme, repeated abuse, and it sits next to Claude's existing ability to end rare abusive chats.1

NVIDIA Omniverse simulation agents

NVIDIA published an Omniverse roundup on October 8. Developers direct GPT-6 Astra and Claude Fable 5 in natural language, and Omniverse libraries supply GPU physics, RTX rendering, and sensor simulation.

Frank DeLise used Astra to turn a SimReady warehouse and humanoid into a first- and third-person interactive simulator. Doyub Kim built Zero to Alpamayo, a reusable Market Street autonomous-vehicle testbed with Omniverse RTX sensors and a Cosmos3-Nano weather and lighting sweep. Ashley Reid ran Astra and Claude Fable 5 against recorded camera and LiDAR. NVIDIA says the agents created two digital twins from scratch and improved two existing ones over about three days.

Tae Kim's Robo Olympics used sports video and Astra to run simulated Unitree G1 humanoids through sports movements. NVIDIA reports one hurdle cleared in 64 of 100 trials. Other internal projects include a wrench designed in Onshape and tested in Isaac Sim, an ISS OpenUSD scene in the browser with telemetry, and stereo captures turned into an editable OpenUSD studio for collision tests. The models are the directors, and the physics stay on NVIDIA's stack.2

Google Cloud MCP Toolbox Java SDK 1.0

Google Cloud shipped MCP Toolbox Java SDK 1.0 last week. It is a stable, backwards-compatible client so Java and Spring shops can give agents type-safe tools against AlloyDB and similar data sources instead of wiring a custom integration for every model.

The useful part is bound-parameter pruning: tenant IDs and other server-owned fields are stripped from the tool definition the model sees, so the agent cannot rewrite them. Authentication is decoupled, with Google OIDC via Application Default Credentials in the box. The SDK also adds a transport abstraction, default tool parameters, session tracking, and a warning when credentials would travel over plaintext HTTP.

The sample app is Cymbal Transit, a fictitious intercity bus network. An AlloyDB concierge answers a pet-policy question, looks up a schedule, and books a seat through parameterized tools rather than raw SQL from the model.3

ChatGPT for Teens College Planner

OpenAI posted a ChatGPT for Teens update on October 7. College Planner is on the way for US students in grades 10 through 12 who are applying to four-year colleges, and it is not generally available yet. The tool is meant to gather requirements, deadlines, tasks, and financial-aid steps for the schools on a student's list. Two-year colleges, technical colleges, trade schools, and more countries are planned later.

OpenAI is supporting College Advising Corps so advisers can test how AI helps students through applications. It is also supporting Boston Children's Hospital Digital Wellness Lab Student Advisory Council for three years. The 2026-27 cohort is about 22 students across digital wellbeing and AI-chatbot tracks.

Study extras include multi-photo notes combined into a PDF on iOS, plus flashcards and quizzes saved in a teen's Library. OpenAI reports about 2.7 million more learning-related messages from teens who had the Teens experience than from those who did not, nearly 1.2 million teens using Learning Visualizations in one week, and average use under 15 minutes a day. Those counts are the company's.4