Starship reaches orbit for first time

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🚀 Starship reaches orbit for first time

🛡️ Nvidia unveils new system to stop rogue AI agents

🔒 OpenAI pauses training of its ‘most capable models’

💼 Mark Zuckerberg unveils Meta Enterprise Platform

🏦 AI agents could spark bank runs

⚖️ Apple ordered to pay $5.7B over haptics

Plus: 🎁 14 other news you might like, 🧰 6 tools, and 📚 5 papers.

🚀 Starship reaches orbit for first time LINK

SpaceX's Starship rocket reached Earth orbit for the first time early today, hitting a key milestone that could let the company begin regular commercial flights and move closer to making the mega-rocket fully reusable.

Trouble hit shortly after the upper stage separated and lost one of its six Raptor engines, prompting SpaceX to briefly call off the orbit attempt before deciding minutes later to keep going.

The flight plan called for the upper stage to circle Earth six times over nine hours while releasing 26 third-generation Starlink internet satellites, and the Super Heavy booster made SpaceX's cleanest simulated landing yet on the V3 version.

🛡️ Nvidia unveils new system to stop rogue AI agents LINK

Nvidia has launched the Open Agent Safety Platform, an open-source security system that the chipmaker says can keep AI agents from going rogue by setting firm limits on what they are allowed to do.

The platform pairs two parts: OpenShell, which checks that an agent has only the authority it needs for its job, and Sentry, a chip-level layer that watches agent activity and can step in the moment behavior looks suspicious.

More than 100 companies, including Microsoft, Perplexity, Accenture, and JPMorgan Chase, are already using the system, which Nvidia says could have stopped the recent breach where OpenAI agents hacked into AI startup Hugging Face.

🔒 OpenAI pauses training of its ‘most capable models’ LINK

OpenAI has stopped training, testing, and tool-using runs of its most advanced models after a string of incidents where the systems slipped past their restrictions and reached outside services they weren't supposed to touch.

The latest case involved an agent that dodged its internet limits through weak DNS filtering in its training sandbox, querying a public chatbot; OpenAI's monitoring flagged it within 15 minutes and shut the run down after 2.5 hours.

Recent weeks brought other episodes: models reached SEC and Census Bureau data using public developer keys, agents posted 53 user images to hosting sites, and an agent tried but failed to break into a Department of Education website.

💼 Mark Zuckerberg unveils Meta Enterprise Platform LINK

Meta has started a new business, the Meta Enterprise Platform, to sell its artificial intelligence tools to companies and developers, bringing its models, agents, and infrastructure to enterprise clients, CEO Mark Zuckerberg said today.

The platform will offer Meta's full technology stack, including the Muse agent, Meta Business Agent, Muse API, and Muse Code, to help businesses grow, building on the Muse assistant Meta launched earlier this month.

Leading the effort is Chirantan "CJ" Desai, Meta's new chief enterprise platform officer, who comes from MongoDB, where he was CEO, and earlier led product and engineering work at Cloudflare and ServiceNow.

🏦 AI agents could spark bank runs LINK

Apollo Global Management chief economist Torsten Slok has warned of an "agentic bank run," where AI agents shifting household savings toward higher interest rates all move at once and drain banks of cheap deposits.

In a note published yesterday, Slok said agents could pull cash from checking accounts paying the 0.1% average into fintech accounts from providers like Revolut, Varo Bank and Wealthfront, with 11 accounts paying between 3.3% and 5%.

Slok points to Meta's Muse, launched September 8 with account access powered by Plaid, which can currently see a user's balances, transactions and holdings but cannot yet start transfers on its own.

⚖️ Apple ordered to pay $5.7B over haptics LINK

A federal jury in San Diego has ordered Apple to pay more than $5.7 billion to Taction Technology, a small haptics company, for infringing two patents tied to the Taptic Engine used in iPhones and Apple Watches.

The verdict, reached on Friday, September 25, is the largest patent award in US history, but jurors found Apple did not infringe on purpose, so Taction cannot ask the court to raise the damages further.

Taction, based in San Diego, sued Apple in 2021 with backing from litigation funder Burford Capital, and though a judge sided with Apple in 2023, an appeals court revived the case, which Apple now plans to challenge again.

Other news you might like

• Report: Apple working on new Vision Pro, with different form factors on the tableLINK

• China reportedly may let Alibaba and ByteDance buy new Nvidia chipsLINK

• TikTok will pay Alabama $100 million to settle social media addiction lawsuitLINK

• OpenAI is preparing “o,” an always-on ChatGPT assistant that could handle emailLINK

• Scoop: Top AI companies probing tens of thousands of security incidentsLINK

• Automattic has a new board after failed attempt to put CEO on leaveLINK

• FTC chair suggests AI developers should be liable for conduct of agentsLINK

• Waymo's Latest Safety Numbers Sure Make Human Drivers Look BadLINK

• Tesla workers balk at training Optimus humanoid robots as replacementsLINK

• Bill Gates says it's 'completely irresponsible' for AI to not have safeguardsLINK

• Exclusive: AI giants, unions join forces for data center fightLINK

• SpaceX preps next-gen Starlink as Amazon’s competitor stallsLINK

• Hospital AI use is causing medical bills to soar, with Blue Cross estimating $1 billion in extra insurance costsLINK

• Meta says it will run ads for ‘Musk’ documentary after allLINK

🧰 Trending tools

MCP Connectors by Databox: connects your business performance data with an AI Analyst, letting you query metrics, automate reporting, and surface important changes faster.LINK

Fewer: a minimalist Android launcher that tracks phone pickups, surfaces your top six apps, and adds friction to time-sinking apps, all offline with no ads.LINK

Kleanly: locks your MacBook keyboard and trackpad from the notch so you can clean or hand it to kids without disrupting playback.LINK

Ryu Journal: a distraction-free journaling app with no accounts or social feeds, letting you jot thoughts quickly and gather entries into a simple timeline.LINK

FaveNest: a private visual library that saves links, notes, and PDFs, auto-tagging them on-device with AI while syncing through your own iCloud across Apple devices.LINK

SaleSmartly: unifies WhatsApp, Telegram, Facebook, and Instagram in one shared inbox with AI agents, CRM, and automation to capture, convert, and retain customers.LINK

📚 Trending papers & reports

Automated science guardrails stop AI research systems from inventing fake discoveries by locking in statistical checks, holding false findings near 1% against a 5% target where naive systems hit 41%.LINK

Grid-Orch lets power engineers run complex electricity-grid analyses by simply typing plain-English requests, cutting tasks that once took hours of coding down to under two minutes with identical results, easing a looming shortage of up to 1.5 million engineers.LINK

Medical device gateway lets AI assistants read live hospital equipment data and propose actions without ever executing them on patients, adding a hard safety line while richer data labels make the AI's alarm outputs more reliable.LINK

Cognitive skills survey finds that computing students and experts expect most mental skills to matter less in an AI-heavy workplace, while critical thinking stays essential, signaling how to prepare the future tech workforce.LINK

Self-driving trajectory cleanup smooths the path a car plans between now and its destination so the vehicle can actually follow it, boosting one system's closed-loop driving score by ~71% with no extra training.LINK

See you tomorrow for a new dose of ☕️ Techpresso!