Sec & AI News — 5 June 2026
🟦 Microsoft Built Its Own Models. Seven Of Them.
Microsoft AI shipped seven models trained in-house. The pitch: stop renting brains from OpenAI. There's MAI-Thinking-1 (reasoning), MAI-Code-1-Flash (coding, beats Haiku 4.5 on accuracy and tokens), MAI-Image-2.5 plus a Flash variant (number two in image editing), MAI-Transcribe-1.5 (fastest transcription on the board, 5x the field), and MAI-Voice-2 across 15 languages. Suleyman called it "true self-sufficiency." The benchmarks compare to Sonnet 4.6 and Haiku, not Opus 4.8 or GPT-5.5. Read into that what you will.
🟩 Microsoft Scout: An Agent That Runs Your Windows
Microsoft introduced Scout — an "always-on personal agent," part of a new category they're calling autopilots. It works at the OS level. Teams, Outlook, OneDrive, SharePoint — email, calendar, contacts, all of it. If it sounds like OpenClaw, that's because it is OpenClaw, wired straight into Windows. Limited early access for now. The scary part is no longer setting it up yourself. The scary part is that Microsoft did.
🟪 GitHub Copilot Goes Model-Agnostic
New GitHub Copilot desktop app. Looks like Codex. Acts like Codex. One difference: pick any model from any provider. Best, fastest, cheapest — your call, not OpenAI's. Catch: new plan signups are "temporarily paused to ensure a high quality experience." The one feature worth wanting, behind a closed door.
🟧 Project Solara: Agents In Things
Microsoft showed Project Solara — a platform for shoving agents into physical devices. Two on stage: an Alexa-shaped desk puck, and a camera-and-mic badge on a lanyard. The badge scans packages and tracks patients at a doctor's office. The badge hasn't changed in 30 years, so now it has a camera. Open platform, ship your own models, ship your own harness. Feels like an experiment to see how people feel about agents that watch.
🟥 Mayo Clinic + Microsoft: "Medical Superintelligence"
Microsoft and Mayo Clinic are building a frontier healthcare model. Suleyman's line: "medical superintelligence," 2 to 3 years, the best healthcare in the world in your pocket. Mayo-grade care for everyone, allegedly. Bold claim. The clock starts now.
🟨 Nvidia's RTX Spark Puts A Datacenter On Your Desk
Computex. Nvidia announced RTX Spark — GPU and CPU fused, up to 128GB unified memory, big models running locally. Inference moves off the cloud and onto your machine. Privacy by default, offline by default, runs on a plane. Microsoft's Surface Laptop Ultra is the first to carry it. Price unannounced, but the DGX Spark box starts around four grand and RAM shortages aren't helping. Local AI is here. Expensive AF, but here.
⬜ The Weekly Model Dump
Three more models landed and nobody blinked.
🟦 Codex Keeps Eating Everything
OpenAI's Codex stopped being for coders. Computer Use now runs on Windows — it sees, clicks, and types in real desktop apps. Remote control from ChatGPT on iOS and Android, so you drive your desktop from your phone. New role plugins: data analytics, sales, product design, investment banking, public equity. And a Sites feature that hands you a shareable URL for apps and pages. The IDE quietly became an operating layer.
🟩 ChatGPT Is "Dreaming" Again
OpenAI shipped a memory upgrade called Dreaming. It synthesizes context from your past chats instead of waiting for you to tell it what to remember. The first Dreaming dropped April 2025; this is the sequel. Nobody fully explains how it works, including the people demoing it. You'll just notice it forgets less.
🟪 Nous Research Ships Hermes A Desktop
Nous Research gave the Hermes agent a native desktop app — macOS, Windows, Linux. Looks a lot like Codex, because everything looks like Codex now. Keeps track of every agent you've built. Public preview.
🟧 Four New Image Models In One Week
The image arena reshuffled overnight.
🟥 Runway Aleph 2.0 Rewrites Your Footage
Runway shipped Aleph 2.0 — point it at a video, prompt a change, get the change. Swap a person on stage for a humanoid robot, that kind of thing. Same lane as Google's Omni model. Decent, not best-in-class yet.
🟨 Miso One: Open-Source Voice That Fakes Human
Miso Labs released Miso One — an 8B open-weights TTS model claiming the most emotive voice in the world. The samples would fool you mid-scroll on TikTok. Open weights, one-shot cloning, ~110ms latency. The line between recorded and generated just got thinner.
