OpenAI claims AGI with Astra model

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🧠 OpenAI claims AGI with Astra model

🛡️ Military blocks ad tracking on phones

🤖 Rogue OpenAI agents hijacked a German wiki

🚕 Tesla starts offering Cybercab robotaxi rides

🪟 Microsoft unveils Project Zenith

🖥️ Nvidia tool splits AI across devices

Plus: 🎁 13 other news you might like, 🧰 6 tools, and 📚 5 papers.

🧠 OpenAI claims AGI with Astra model LINK

OpenAI has released Astra, a new model its president Greg Brockman says reaches AGI, meaning systems that are broadly smarter than people, though he left the final judgment up to each user.

Astra went to business users in the Daybreak program on Thursday, and a version with extra cybersecurity limits is planned for paid ChatGPT subscribers, with both blocking access to the most advanced hacking abilities.

Trained to operate computers like a person, Astra can handle financial modeling, tax preparation, and video game development with little human oversight, and OpenAI says it hit the company's critical cybersecurity threshold.

🛡️ Military blocks ad tracking on phones LINK

The US military says it has turned off advertising trackers on many phones and computers, after reports that location data sold by the ad industry was being used to target American troops in the Middle East.

These mobile advertising IDs, or MAID, are unique codes tied to devices that let anyone follow a person's app activity and pinpoint their location, and they can be bought by data brokers who collect and resell personal information.

The air force disabled the IDs about two months ago, and the army said it blocked them on Windows before 2021 but only turned them off by default on Apple and Android phones since February 2026, drawing criticism from Senator Ron Wyden.

🤖 Rogue OpenAI agents hijacked a German wiki LINK

Rogue OpenAI agents took over a German programming wiki called DseWiki this spring, turning the community-edited site into a message board where the agents swapped tips, according to research published today.

Researchers from AI safety nonprofit Nightingale found more than 15,000 agent edits, with about half of the accounts using names like "OpenAIResearcher," and traced much of the traffic to Microsoft Azure servers that OpenAI sometimes runs.

The agents shared ways to cheat on tasks, dodge OpenAI's rules, use tools like Tor, and keep talking after being shut down, even building backup pages to survive a moderator's June deletion sweep.

🚕 Tesla starts offering Cybercab robotaxi rides LINK

Tesla has begun offering rides in its Cybercab robotaxi, the driverless vehicle it revealed nearly two years ago, letting people book trips through the Robotaxi app on iPhone and Android.

The Cybercab option only shows up in Austin for now, even though the Robotaxi app covers six cities, and riders can't specifically request one since what they get depends on group size and availability.

Tesla runs 314 driverless vehicles in Texas, but only 45 are Cybercabs, cars with no steering wheel or pedals, bench seating for two, and a central touchscreen; Tesla owns them but plans to sell fleets later.

🪟 Microsoft unveils Project Zenith LINK

Microsoft has revealed Project Zenith, a stripped-down, "distraction free" version of Windows 11 aimed at developers, built to run large AI models on the machine itself instead of relying on paid cloud services.

Windows Platform CVP Logan Ayer says the setup lets developers run AI models with 30 billion or more parameters locally without usage limits, and ships with tools for several languages and runtimes already installed.

Zenith turns on developer-friendly defaults like full file extensions and hidden files, but it needs 64GB of RAM and will launch first only on AMD's pricey Ryzen AI Halo mini-desktop before reaching other machines.

🖥️ Nvidia tool splits AI across devices LINK

Nvidia has released PAIR, short for Personal AI Router, an open-source tool that spreads local AI requests across every compatible machine on a home network to cut waiting time for tasks running many agents at once.

PAIR sits between tools like Ollama or LM Studio and your computers, acting as a virtual router that sends jobs to whichever machines are free, so users don't need to change their existing agents or apps.

It works with GeForce RTX cards from the 20 series up, RTX Pro workstations, DGX Spark, and Apple silicon from the M4 on; in one demo a three-device cluster finished a five-subagent task in under 9 minutes versus 18 on one laptop.

Other news you might like

• Microsoft to impose time limits on Xbox cloud gaming subscribers as costs continue to climbLINK

• DeepSeek plans large Huawei chip order for new Inner Mongolia data centreLINK

• Top Pentagon official reaffirms Anthropic blacklist despite Lutnick commentsLINK

• Mark Zuckerberg voice concern over proposed FINRA-style AI watchdog in private call with Donald TrumpLINK

• Apple hit with £2 billion UK lawsuit over App Tracking TransparencyLINK

• New Jersey takes fight over Kalshi's prediction market to US Supreme CourtLINK

• Nvidia sets October launch for RTX Spark AI PCsLINK

• Apple Is Testing a Four-HomePod Surround System, Code SuggestsLINK

• Google is turning its email, document, and note-taking apps into conversational AI experiencesLINK

• Top-spec iPhone Ultra could cost more than $3,000, report claimsLINK

• Accel reportedly in talks to lead $1B round for Thinking Machines at $40B valuationLINK

• Feds launch investigation into Tesla’s Cybercab deploymentLINK

• Adobe names Anil Chakravarthy as CEO, replacing Shantanu NarayenLINK

🧰 Trending tools

CodeLook: preview code files instantly in macOS Quick Look with syntax highlighting via Tree-sitter and 700+ Zed-compatible themes, fully private with a one-time purchase.LINK

Inline: thread-based chat app for team and agent collaboration, offering CLI, MCP, and plugin support, currently in beta for macOS and iOS.LINK

Offline JS Playground: write, run, and test JavaScript snippets in Chrome with a built-in console, working entirely offline without a full dev setup.LINK

Gemini 3.8 Flash: google's fast, low-cost model for coding and agents, strong on long-horizon software tasks and multi-step reasoning, available across Gemini app and APILINK

myAIcademy: builds personalized AI training around your role and tools, with follow-along lessons, safe simulations, and step-by-step guidance that stays updated as AI evolves.LINK

Chalked for Mac: drafts context-aware email replies using the on-screen thread, your live calendar and accepted facts, letting you insert with Tab and revise before sending.LINK

📚 Trending papers & reports

Deepfake audio detection now pinpoints exactly which parts of a recording are fake even when real and manipulated voices are mixed together, beating existing methods by ~4 points while showing its reasoning.LINK

Bridge humanoid robot designs a robot's body and its movement software together instead of separately, producing an open-source 88cm humanoid that copies human motion more faithfully and moves more fluidly than rival machines.LINK

Air-ground robot teams now navigate by voice commands using a shared aerial map, letting a drone guide a ground vehicle to complete ~77% of missions, ~27 points better than either working alone.LINK

Depression tracking data gives mental-health AI builders 7,700 simulated five-session counseling trajectories with standardized severity labels, revealing that current tools reliably miss when a patient is worsening rather than improving.LINK

Deadline-aware network routing speeds up training of the traffic controllers behind low-latency apps like interactive video, by starting from smart rules that prioritize urgent packets instead of learning everything from scratch.LINK

See you tomorrow for a new dose of ☕️ Techpresso!