OpenAI shipped GPT-Live, GPT-5.6, and ChatGPT Work in two days.

The feature that matters most got almost no airtime.

ChatGPT Work is the headline product. Give it a goal and it can plan the approach, break the work into steps, delegate parts of the project to sub-agents, and work through longer, more involved tasks across connected apps and files.

SharePoint. Teams. Google Drive. Email. Calendars. CRMs.

More than 1,400 plugins.

But in the new desktop app, Work can also be given Computer Use.

With the required OS, application, and folder permissions, it can work with local files, applications, browsers, and other tools including creating and changing files directly on the machine.

Which brings me to legal.

If you run technology at a law firm and have lawyers are using unmanaged or shadow AI in your environment this deserves attention.

A user who enables Computer Use and approves the required permissions can give an autonomous agent meaningful access to a laptop that may connect to content from the DMS and contain confidential and potentially privileged email.

It is not just one toggle.

But it may be only a handful of permission prompts that users click through without fully understanding the consequences.

Rule 1.6 does not care how easy the prompts were to approve.

Agents are clearly where knowledge work is heading, and the firms that figure out governed adoption will pull further ahead of the ones that freeze.

But firms should decide who gets Computer Use, restrict plugins and applications by role, configure approval requirements, limit local-file and network access, set usage and spending controls, and tell their people what changed before they discover it on their own.

Because the announcements will not do it for them.

That was kind of the point.