Agents everywhere: helping, hacking, leaking, and costing a trillion

Meta wants its Muse AI agents to run big parts of our lives. In Mark Zuckerberg's vision, your agent shops, works, and talks to other people's agents for you.

But this week was rough. Security researchers at Claroc found that public Muse conversation links were easy to find. They saw a marketing company's internal revenue review and a small business owner's loan application. Meta paused sharing, and has now turned it back on.

Also, Meta changed its business terms to remove words suggesting it could freely train AI on news sites. Legal experts told The Verge this is not enough. Other language could still let Meta scrape them.

OpenAI is preparing to go public. Before that, it is working on a new raise of $30 billion at a $1.25 trillion valuation.

If it closes, OpenAI would be the most valuable private company ever. In China, DeepSeek reportedly raised $15 billion, one of the country's largest rounds. That came only days after the US ordered federal agencies to remove DeepSeek's technology for national security reasons.

The chip race is also eating money. The Verge reports SpaceX is about to borrow another $40 billion to spend on Nvidia chips. These deals are reported, not all final.

Healthcare is the promise AI leaders repeat the most. Now Microsoft AI CEO Mustafa Suleyman says "it is finally here."

Microsoft's diagnostic system beat generalist doctors on hard cases, in a study published in Science. Microsoft claims it can diagnose conditions up to four times more accurately than human doctors. Suleyman talks about chasing "medical superintelligence."

But TechCrunch notes a key detail. Most of the improvement seems to come from the AI models underneath, not from Microsoft's own system. Also on health: inclusionAI released Sante, a free open-weight medical model (you can download and run it yourself).

A new ABC News/Washington Post/Ipsos poll says most Americans think AI is bad for society. Most also oppose building more data centers.

In Sweden, public radio reports that government offices will ban ChatGPT and Claude. The report does not give all the reasons in our sources.

Energy is part of the debate. A Swedish AI company, Berget, says the same AI request emits 8.5 times more CO2 in Texas than in Sweden, because of how the electricity is made. That is a company press release, so take it as one view.

If agents are going to act for us, they need safe limits. Two new tools try to give them that.

NVIDIA released OpenShell on GitHub. It is a "runtime", a protected space where autonomous AI agents can run safely and privately, even in large fleets. It got 427 points on Hacker News. It is mostly for developers and companies running many agents.

1Password launched a way for agents to log into your personal accounts without ever seeing your password. And Anthropic donated its Agent Skills standard to the Agentic AI Foundation under the Linux Foundation, to keep it open and vendor-neutral.

Canva, the design app many small businesses use, now has its own AI image model. It is called Canva Craft.

It works in two stages. First it plans the design, then it makes the image. Canva says this gives more precision, consistency, and control. It competes directly with image models from OpenAI, Microsoft, and Google.

Craft is available to all Canva users. Canva promises better, more customizable results at lower cost.