Microsoft is adding a feature to 365 Copilot that lets AI agents handle web tasks on their own, pushing automation in knowledge work a step further.
The new “Researcher with Computer Use” tool acts like an autonomous browser agent, similar to the ChatGPT agent. It can click, navigate, and run code to automate complex searches, analysis, and report generation.
Each session runs inside a sandboxed virtual machine (VM) created with Windows 365. This isolated cloud computer includes its own browser, terminal, and text editor, and remains completely separate from both the company network and the user’s device. Microsoft says login data isn't stored or transmitted.
The agent can access information behind logins, such as paywalled articles or company databases, if the user or admin approves. When access is needed, the system asks for confirmation, letting the user step in to log in or approve specific actions.
Microsoft gives an example where the agent downloads a World Bank dataset from the terminal and uses Python to analyze national savings rates.
Admins manage access, but security risks persist
By default, the tool blocks company data like emails, SharePoint, and meetings. Users can share specific data sources, but administrators decide which user groups get access, what data can be combined, and which websites are allowed. Microsoft says all sandbox activity can be audited.
Even with these safeguards, autonomous AI systems still pose security challenges, especially when interacting with external content. Studies continue to warn about the risks of letting AI agents operate freely on the open web.
