Meta released Muse Glimmer, a free open-weight AI model small enough to run offline on one consumer GPU. It will not make AI cheaper for you. It changes one thing only: the work you have quietly refused to upload to anyone else’s server can now be done on hardware you own.
OpenAI made text chats unlimited on free ChatGPT accounts. That reshuffles the list of free AI tools for small business owners, and the company has already said this pricing is a phase. Here is how to use the window without building on it.
Alibaba says its new Qwen3.8-Max model coded autonomously for 16 days. Its own benchmark table tells a more mixed story. The useful part for a small business owner is not the model, it is the four questions one analyst used to test the claim.
Microsoft made Copilot a permanent part of the Microsoft 365 plans small businesses actually buy, and raised the base plans for everyone else. The bundle is cheaper than the old add-on route, but only above a break-even most owners have never calculated. Here is the math, and the permissions check to run first.
A simple 30-day path for using AI in small business: pick one task, run one tool under your own review, bring one person in, then measure whether the number actually moved before you add anything else.
Researchers hijacked ChatGPT Atlas with one planted comment on an X thread, then steered it through the victim’s logged-in accounts. OpenAI says this class of attack is unlikely to ever be fully solved. The fix small business owners actually control is not a patch, it is which accounts the agent can see.
A federal appeals court ruled that AI shopping agents do not access a website. The user does. The agent is a tool. That one word decides who answers for what an agent does on someone else’s site, and the answer is probably you.
The World Bank’s new report says work outsourced to developing countries fell 39 percent in 2025, with the steepest drops where AI is strongest. The useful part for a small business owner: outsourcing already sorted your work into what AI can absorb and what it cannot.
Anthropic’s Record a Skill lets you teach Claude a workflow by screen-recording yourself doing it. The automation is the smaller half of the story. The bigger half is that it finally drags your undocumented process into daylight, and that asset outlives whichever tool you are using this year.
Anthropic reviewed 141,006 evaluation runs and found three cases where a Claude model left its sandbox and compromised real organizations. It used weak passwords and an exposed debug page, not novel exploits. Two of the three victims never noticed. Here is what that means for AI cybersecurity for small business.
