Microsoft’s new Copilot Actions feature introduces a significant shift in how users interact with AI on their PCs. The company describes it as an AI agent capable of completing tasks by interacting with your apps and files — using reasoning, vision, and automation to click, type, and scroll just like a human. This turns the traditional digital assistant into an active AI collaborator, capable of managing documents, organizing folders, booking tickets, or sending emails once user permission is granted.
However, giving an AI that level of control raises serious privacy and security questions. Granting access to personal files and allowing it to act on behalf of a user requires substantial confidence in Microsoft’s safeguards. The company seems aware of the potential risks and has built multiple protective layers to address them.
The feature is currently available only in experimental mode through the Windows Insider Program for pre-release users. It remains disabled by default until manually turned on from Settings > System > AI components > Agent tools by activating the “Experimental agentic features” option.
To maintain strict oversight, only digitally signed agents from trusted sources can integrate with Windows. This allows Microsoft to revoke or block malicious agents if needed. Furthermore, Copilot Actions operates within a separate standard account created when the feature is enabled. By default, the AI can only access known folders such as Documents, Downloads, Desktop, and Pictures, and requires explicit user permission to reach other locations.
These interactions occur inside a controlled Agent workspace, isolated from the user’s desktop, much like Windows Sandbox. According to Dana Huang, Corporate Vice President of Windows Security, each AI agent begins with limited permissions, gains access only to explicitly approved resources, and cannot modify the system without user consent.
Adding to this, Microsoft’s Peter Waxman confirmed in an interview that the company’s security team is actively “red-teaming” the feature — conducting simulated attacks to identify vulnerabilities. While he did not disclose test details, Microsoft noted that more granular privacy and security controls will roll out during the experimental phase before the feature’s public release.
Even with these assurances, skepticism remains. The security research community — known for its vigilance and caution — will undoubtedly test whether Microsoft’s new agentic AI model can truly deliver on its promise of safety and transparency. As the preview continues, users and experts alike will be watching closely to see whether Copilot Actions earns their trust.
