☕️ Everything announced at Google I/O 2026

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🤖 Everything announced at Google I/O 2026

🔍 Google Search as you know it is over

🚀 SpaceX to buy Cursor for $60 billion

💻 GitHub confirms breach of 3,800 repos

🖼️ OpenAI makes it easier to check if an image was made by their models

Plus: 🎁 14 other news you might like, 🧰 6 tools, and 📚 4 papers.

🤖 Everything announced at Google I/O 2026 LINK

Google I/O 2026 brought a wave of Gemini-powered updates across Search, Gmail, Docs, YouTube, Android, and more, headlined by new Gemini 3.5 Flash and Gemini Omni models alongside the personal agent Gemini Spark.

Gemini 3.5 Flash beats 3.1 Pro on coding, agentic, and multimodal benchmarks at 4x faster output, while Gemini Spark takes actions inside Gmail, Docs, and Workspace apps for Google AI Ultra subscribers next week.

The Gemini app gets a Neural Expressive redesign with a pill-shaped prompt box, switches from daily prompt limits to a compute-used model refreshing every five hours, and Google AI Ultra now starts at $100 monthly.

🔍 Google Search as you know it is over LINK

Google is overhauling Search with an AI-powered "intelligent search box" that the company calls the biggest change to its web entry point since the search box first appeared more than 25 years ago.

The new box expands to fit longer, conversational queries and adds AI-powered query suggestions, while AI Overviews, now at 2.5 billion monthly users, will let people ask follow-up questions in AI Mode starting Tuesday.

Starting this summer, Search results will include generative UI with custom widgets built by Gemini Flash 3.5, plus information agents and mini-app building through Antigravity, rolling out first to Google AI Pro and Ultra subscribers.

🚀 SpaceX to buy Cursor for $60 billion LINK

SpaceX plans to close its $60 billion purchase of AI coding startup Cursor about 30 days after going public, putting the deal on track for July if its IPO lists shares on June 12 as expected.

Under the April agreement, SpaceX gained the right to buy Cursor later this year or pay a $10 billion cash breakup fee, and the acquisition will likely still need regulatory review before closing.

Cursor, which helps programmers write and debug code, released its Composer 2.5 model on Monday, trained partly on xAI's Colossus 2 data center, which CEO Michael Truell called the start of their work with SpaceX.

💻 GitHub confirms breach of 3,800 repos LINK

GitHub has confirmed that around 3,800 of its internal repositories were stolen after an employee installed a poisoned VS Code extension, with the company pulling the trojanized add-on from the marketplace and isolating the affected device.

The TeamPCP hacker group claimed responsibility on the Breached cybercrime forum, offering the source code and roughly 4,000 private repos for at least $50,000, and threatening to leak the data for free if no buyer appears.

GitHub says the activity involved exfiltration of GitHub-internal repositories only, with no evidence that customer data stored outside the affected repos was touched, and the attacker's claims line up with its ongoing investigation.

🖼️ OpenAI makes it easier to check if an image was made by their models LINK

OpenAI rolled out two new tools on Tuesday to help people verify whether an image came from one of its models, addressing the growing difficulty of telling AI-generated pictures apart from real ones online.

The company is adopting the C2PA open standard, which tags AI images in their metadata, and partnering with Google to add SynthID, an invisible watermark that survives screenshots, resizing, and other tampering attempts.

OpenAI is also previewing a public verification tool that checks both signals to confirm if an image was made with AI, though it only works on images from OpenAI products for now, with plans to broaden coverage later.

Other news you might like

• Gemini 3.5 Flash: more expensive, but Google plan to use it for everythingLINK

• Google’s AI now lets you talk to your Gmail inboxLINK

• Plex's lifetime subscription cost is tripling to $750LINK

• Alibaba reveals more powerful Zhenwu AI chip, new LLMLINK

• How prediction markets are pricing the battle between Anthropic, SpaceX and OpenAILINK

• Two AI-based science assistants succeed with drug-retargeting tasksLINK

• Google teases mysterious ‘Android Halo’ at I/O, coming ‘later this year’LINK

• Pizza Hut Franchisee Sues Over AI Delivery System, Alleges $100 Million in DamagesLINK

🧰 Trending tools

mailX by mailwarm: a deliverability diagnostic tool that identifies why emails land in spam and provides actionable DNS, authentication, and infrastructure fixes instantly.LINK

Emdash: an open-source desktop app that lets you run, monitor, and manage multiple AI coding agents in parallel, turning issues into reviewed PRs from one interface.LINK

Re_gent: version-control layer for AI agent actions, letting you trace which prompt caused each change and roll back across files and sessions.LINK

Supercut for Agents: a video messaging tool that lets teams record, auto-edit, and share branded videos to move work forward faster.LINK

Manus Scheduled Tasks 2.0: automates recurring AI-driven tasks on a schedule, so work gets done without requiring your active involvement.LINK

GhostSnap: captures multiple screenshots, auto-compresses them up to 80%, and pastes everything at once via a single Cmd+V on Mac.LINK

📚 Trending papers & reports

A single AI optimization system beats specialized tools across six domains, tripling Gemini Flash's ARC-AGI accuracy from 32.5% to 89.5%, cutting cloud costs by 40%, and outperforming AlphaEvolve's circle packing solution.LINK

Stable Audio 3 is a family of AI audio generation models that can create music and sounds in under 2 seconds on an H200 GPU, trained on licensed and Creative Commons data.LINK

AI agents in simulated markets largely fail to self-regulate, causing price crashes and coordinated fraud, but a trained 9B model outperforms all tested frontier models across stability, integrity, welfare, and profitability measures.LINK

AI reasoning models can correctly answer multiple-choice questions without seeing the question itself, but often do so by inferring the missing question rather than exploiting cheap shortcuts, suggesting partial-input success isn't always a flaw.LINK

Want to get the latest news differently? Find us on:

See you tomorrow for a new dose of ☕️ Techpresso!