Safety and control

An agent that can use a computer acts with real permissions. These are the controls that make that safe.

Least privilege

Give the agent its own account with only the access the task needs.

Human approval

Require a person to approve payments, deletions, external messages and other sensitive actions.

Audit trail

Record screens and actions so every run can be reviewed.

Isolation

Run agents in a separate virtual machine or browser profile, away from personal data they do not need.

Guarding against prompt injection

Text on a web page or in a document can try to give an agent new instructions. Treat on-screen content as data, not commands, and keep humans in the loop for anything consequential.

Public frameworks such as the EU AI Act (Regulation (EU) 2024/1689) and the NIST AI Risk Management Framework set out broader expectations for trustworthy AI.