Computer Use

Most AI acts through APIs: clean, structured, permissioned. Computer use takes the general path instead — the model looks at a screenshot, decides where to click and what to type, and acts through the graphical interface.

The appeal is universality. It works with software that has no API, no integration and no vendor cooperation, which describes an enormous amount of the software real organizations actually run. Nothing else reaches legacy internal systems.

The reliability is not yet there for unattended use. Interfaces vary endlessly, screens change, dialogs appear, and an agent that misidentifies a button does something other than what was intended. It is also slow, since every step is a screenshot and a decision.

The security implications are the sharpest of any AI capability: an agent operating your computer inherits your access to everything on it.

This is potentially the broadest-reach agent capability, because it needs no integration work at all. It is also the one where a prompt injection on a viewed web page translates most directly into damage, since the agent can act on anything the logged-in user can.

Run it in a sandbox or dedicated VM with its own credentials, not on a machine holding your real sessions. Keep a human approving consequential steps, and treat every page it views as untrusted input.