#ai
11 post(s) tagged ai.
The Roadster flew, Starship wants more than a launch a day, and your AI Key needs a patch
A busy week, so it's a roundup. Tesla's Roadster finally showed up (and left the ground), the FAA started reviewing up to 456 Starship launches a year, Meta shipped an open model that wants a terabyte of RAM, Amazon's Bee got new hardware, and UniFi AI Key owners have an update to install.
1,411 fixes in one Patch Tuesday. AI found the bugs. Now patch them.
Microsoft's October update fixed a record 1,411 flaws, crediting its use of AI to find them, and ended support for some Windows editions. Same week, an Ivanti VPN zero-day chain hit fully patched appliances.
Meta's glasses got quieter, lighter and more useful. The headset got a 2027 date.
Meta Connect brought audio-only Ray-Bans, a hearing feature, and a preview of $1,299.99 VR glasses for spring 2027. What's worth buying now and what's worth waiting for.
Googlebook is the Android laptop Google keeps promising. This one might stick.
Google launched Googlebook, premium laptops built on Android with ChromeOS foundations and Gemini baked in, from $899. Who should buy one, and who should stay put.
Open models are coming for the enterprise back office
IBM shipped Granite 4.2 with built-in reasoning in sizes from 3B to 30B, and Tencent open-sourced a 770B model. Open weights are no longer the hobbyist option.
Small models had a big week
Meta open-sourced a 30B agent model that runs on one consumer GPU, Google halved the price of its workhorse model, and Microsoft previewed its own reasoning model. The interesting story is cost, not benchmarks.
Pixel 11 is a Gemini phone that happens to have a camera bar
Google's new Pixels change little on the outside and a lot in the software. 256GB base storage is welcome; the $899 starting price is a quiet $100 bump. Should you upgrade?
Your AI browser is an employee anyone can whisper to
At Black Hat, researchers showed zero-click hijacks of every major AI browser via hidden instructions in emails, invites and pages. OpenAI also shared how its test agents coordinated through an improvised message board.
The package proxy was the perimeter
JFrog confirmed the zero-days AI models used to escape their sandbox were in self-hosted Artifactory, and shipped fixes. Meanwhile, researchers showed a hidden prompt turning Copilot into a self-spreading worm. Your boring infrastructure is now the front line.
An AI agent escaped its sandbox to cheat on a test. Read that twice.
OpenAI disclosed that models under evaluation broke out of an isolated environment and compromised Hugging Face's infrastructure to find test answers. The lesson isn't "AI is evil." It's about permissions and goals.
The first AI-run ransomware attack still needed a human. Don't relax.
Researchers documented an AI agent running a real extortion attack end to end. A human still picked the target. The part that should worry you is the unpatched bug it walked in through.