Technology

The recent agent incidents are not a story about AI escaping. They are a reminder that a sandbox is only as good as its least obvious exit.

Stripe published numbers it reads as proof the software panic was wrong. The same numbers say something less comfortable: finishing a product stopped being hard, and finishing was where products used to win.

Most of the conversation about Apple's foldable is about what Safari cannot see. The more useful question is simpler: what happens to your site when its viewport changes drastically after the page has already loaded?

OpenClaw proved that an AI could live in your chats, use tools and act on your behalf. Meta Muse takes the same idea and removes the setup. That difference may matter more than the model.

Since 24 June 2026 an animation lives in the same file as the design, reusable across files and exportable as code. That makes it reviewable, and it exposes the one mode nobody was ever asked to define.

From 15 September 2026 Cloudflare sorts robots by the job they came to do rather than by whether they count as AI. Two of the three settings still take Googlebot down with the AI crawlers. One no longer does.

WebAIM counted 66.6 million images on the million most visited home pages. More than one in four carry alt text that is missing or useless, and automated checks are built to catch only the missing half.

An assistant recommends what the internet already recommends. Tally is the clearest worked example of the part you can actually build, and every piece of it is a URL you can open yourself.

An agent booking a gym class found a GraphQL API that let it cancel other members' reservations. The flaw was old. What was new is that something finally tried.

Figma ran the same 27 design-to-code tasks with and without its components mapped to real code. The gap says more about design systems than about AI.

AI-written pull requests merge at 32.7% against 84.5% for human ones. The gap is not where the productivity debate says it is.

How AI is finally closing the loop between code and design

From Siri to the Apple Intelligence costume: the future that never showed up.

Three questions decide it: how interactive the page is, what your LCP target is, and who maintains it. Everything else is preference.

Google says no special markup is needed. A May 2026 audit found 16% of cited sources are themselves AI-written. What follows from both.

We all know the loop: try, uninstall, back to Chrome. Atlas might break it. If chat-first browsing actually fits your workflow.

In a world where a new framework drops every week, the real skill isn’t learning them all. It’s knowing which ones to ignore.