AI news for Friday, October 16, 2026: Meta's Muse Spark Open fits on two Macs, ChatGPT works in Chrome

Big models got local and your browser got an agent, so today's job is deciding what each one is allowed to touch.

Open source Models3 min
Meta's Muse Spark Open fits on two Mac Studios
Meta released Muse Spark Open, an 816B-parameter open-weight model it says runs on two M5 Ultra Mac Studios. The size is real. The scorecard and the full license terms are still missing.
Why it matters If you run models on your own hardware, a frontier-size open model now fits in a closet instead of a data center. If you build on it, test it on your own coding tasks and read the Muse Model License first, because Meta has published neither benchmark scores nor plain commercial terms.
SRCMeta
Agents Products3 min
ChatGPT now reads your Chrome tabs and works in them
OpenAI updated its ChatGPT extension for Chrome on Oct. 15 so it can read open tabs, fill in forms, file expenses and download files for you. That puts an agent in the browser where your logged-in accounts already live.
Why it matters An agent that can act on logged-in sites can also be steered by any page it reads. Before you turn on Work mode, give it a separate Chrome profile, keep banking and email out of it, and start with low-stakes chores.
SRCOpenAI (Chrome Web Store)
Policy Safety2 min
FTC demands OpenAI and Anthropic safety records
The FTC sent civil investigative demands to OpenAI and Anthropic seeking documents on safety incidents, privacy policies and safeguards, an agency spokesperson confirmed to CBS News. Two weeks ago this was an inquiry. Now it is compulsory.
Why it matters Your API keeps working today. But when a regulator asks labs how their agents are fenced in, expect your own customers to ask you the same question. If you deploy agents on these models, write down what they can reach, what they log and who reviews them.
SRCCBS News
Creative Open source3 min
Black Forest Labs' FLUX 3 Klein runs on an 8 GB GPU
Black Forest Labs released FLUX 3 Image Klein, a 5.6B-parameter open-weight image model whose INT8 build runs on GeForce RTX cards with 8 GB of VRAM. If you earn under $5 million a year, you can run a serious image model in-house for free.
Why it matters Image generation for products, ads and mockups no longer needs an API bill or a data center GPU. If you're under the $5 million revenue line, you can run and fine-tune it on one gaming card. If you're over it, the license conversation comes first.
SRCBlack Forest Labs
Models Business2 min
Microsoft builds a 2.8T-parameter model of its own
Microsoft AI finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total and 80 billion active parameters, and says it is 4 to 12 times more compute-efficient than DeepSeek's base model. Nothing ships yet. The signal is that Microsoft wants its own frontier model.
Why it matters If you buy AI through Azure, expect a Microsoft-made model on the Foundry menu next to the ones you use now. There is nothing to test today, and the efficiency numbers are Microsoft's own, so wait for the post-trained model and public benchmarks before you plan around it.
SRCMicrosoft AI

Prompt of the day

Write the rules for an AI agent in your browser

› I'm about to let an AI agent work inside my web browser. It can read my open tabs, fill forms, send messages and download files. Help me write a short, practical rulebook before I turn it on.

My situation:
- What I want the agent to do: [e.g. research, tab cleanup, expense reports, CMS data entry]
- Sites I'm logged into most days: [e.g. Gmail, bank, Slack, Shopify admin]
- Who else is affected: [just me / my team of X / my clients]

Give me:
1. Three lists: sites the agent may read, sites it may act on, and sites it must never touch. Explain each choice in one line.
2. The actions that must always wait for my explicit confirmation (sending, paying, deleting, posting), tuned to my sites.
3. How to set up my browser so the agent works in a separate profile or window, step by step.
4. Five warning signs that a web page is trying to give the agent instructions (prompt injection), written so I can spot them myself.
5. A 10-minute weekly review I can run to check what the agent did.

Keep it plain and specific to my answers. Ask me up to three questions first if anything above is missing.

Open in Magai

Works in any AI chat. In Magai, run it against several models at once.

Get every edition

This one arrived in inboxes at 6am ET.

Meta's Muse Spark Open fits on two Mac Studios

ChatGPT now reads your Chrome tabs and works in them

FTC demands OpenAI and Anthropic safety records

Black Forest Labs' FLUX 3 Klein runs on an 8 GB GPU

Microsoft builds a 2.8T-parameter model of its own