A computer-use model is an AI system that interprets a graphical or software interface and takes actions through it, such as clicking buttons, entering text, navigating pages, or operating applications. Unlike a system that only generates answers, it can carry out steps in a workflow.
This matters in cybersecurity because actions can affect real data and systems. If connected to email, files, browsers, or administrative tools, a model’s permissions become part of the organization’s trust boundary. Attackers may also try to steer it with malicious instructions embedded in content it reads. Defenses include limiting access to necessary functions, isolating the agent in a sandbox, logging its actions, and requiring human approval for sensitive operations.