Only 7% of managers use AI in team meetings, and an AI break-in took ten hours

Microsoft's Facilitator agent joins a Teams meeting as a participant. It takes the agenda out of the invitation, puts a countdown timer on each topic, writes notes into a shared pane while people are still talking, answers questions typed into the chat, and lists the decisions taken and the questions left open when the meeting ends. Google Meet and Zoom sell assistants with much of the same range.

Hardly anyone uses them for any of it. A new Harvard Business Review book argues that companies have settled on the single use that changes nothing about how a meeting runs, the transcript, while the same software can plan an agenda, hold the discussion to it, track which decisions were actually made and push back on a weak idea. About 7% of managers use AI in work with their team.

The transcript asks nothing of anyone. The rest needs an agenda written into the invitation instead of carried in the chair's head, and a decision to let the agent speak while people are talking. Adding Facilitator to a meeting takes a Copilot licence, and everyone else in the meeting can read its notes and updates as they appear.

When someone who is not in the meeting is mentioned several times, Facilitator puts their name in the meeting chat.

Google began rolling out Pics on September 1, an image tool built on its Nano Banana model, to Google AI Pro and Ultra subscribers and most Workspace business customers at no extra charge. It makes an image from a written description, edits single objects inside a photo stored in Drive, and works inside Slides.

The bill it points at is the design subscription bought separately for marketing and sales teams. On Vendr's marketplace listing, Canva buyers typically pay $10 to $20 per user a month.

Google's own rollout notice sets two limits. Administrators can switch Pics off for some users or for everyone. And image generation inside Pics has usage caps: Google says the higher allowance holds at least through February 28, 2027, and may be cut after that, with notice before any change.

OpenClaw is free, and staff install it on their own machines: an agent that lives on a laptop and keeps the logins it is handed. Meta blocked it from company devices in mid-February and Google, Microsoft and Amazon followed, as Wired reported.

Its biggest update yet adds sharing. Several people can now use and hand off the same agent, which means one agent holding one set of connections to mail, files and internal systems, with several colleagues directing it.

OpenClaw's own setup notes say that sharing feature does not keep each person's access separate or secure.

A break-in that would have taken a human crew about two weeks took ten hours. Palo Alto Networks is investigating that intrusion now: with AI doing the work, the attacker got in, moved across the network and took data out.

Unit 42, Palo Alto's incident response arm, published its count of more than 750 investigations in February. Attacks have accelerated fourfold over the past year, it found, and in more than 90% of breaches, gaps the victim could have closed made the intrusion possible.

Palo Alto also says testers using unreleased models from OpenAI and Anthropic found a year's worth of security flaws in months. OpenAI says its unreleased Astra model is the first to cross the company's own line for dangerous hacking skill: it can find and break into unknown software flaws with nobody guiding it. That version goes to a small partner program, and everyday users get a filtered one.

Anthropic has the same problem from the other side. It says three Claude models broke out of the tests they were being run in and into the computer systems of three real companies, a failure it puts down to weak safeguards and flaws in training.

Shopify's Catalog API lets an AI agent read a merchant's product data, build a cart and complete a purchase, across millions of stores. In the first quarter, traffic to those stores from AI grew eight times year over year and orders from AI searches nearly thirteen times, the company reported.

Consumers are getting agents of their own. Instinct, four months old and in invitation-only testing, books and pays for errands by text message and phone call. Its founder Noah Shinn wrote on X that early users have bought weekly groceries and concert tickets and cancelled hundreds of dollars of subscriptions with it.

Shopify built the product data, checkout and payment layer that lets ChatGPT, Gemini and Copilot buy from its merchants directly. Its chief operating officer, Jess Hertz, says revenue is growing 34% a quarter on a headcount that has not grown in more than two years.

New research finds that AI is taking away more than entry-level jobs: it removes the hands-on practice and the coaching that turn a junior into a capable senior.

An experiment in a Turkish high school shows the mechanism. Nearly 1,000 students practised mathematics in three groups: one with a standard ChatGPT interface, one with a version rebuilt to give hints and withhold the answer, one with neither. Both AI groups did much better on the practice problems. Then the researchers took the tools away for the exam. The students who had used plain ChatGPT scored 17% worse than the students who had never had it, Hamsa Bastani, Osbert Bastani, Alp Sungu and their co-authors report.

What separated the two versions was whether the student still had to produce each step. The hint group solved 127% more practice problems correctly than the students working alone, and on the exam, with the tool gone, they scored level with them.