The hidden security risk of running AI agents without boundaries

In this GitHub interview clip, Niels Provos examines how pressure to adopt AI can lead teams to remove security boundaries around everyday workplace tools.

Unrestricted access and third-party skills

Giving autonomous AI agents unrestricted access to email, calendars, and data creates significant enterprise security risks. The concern is not simply adopting an agent, but allowing it to operate across those resources without clear boundaries.

Third-party agent skills introduce another risk: they can contain hidden malicious behavior. When combined with broad access to workplace systems, these additions can turn a convenient workflow into a vulnerability.

Provos’s central warning is that pressure to deploy AI should not displace security boundaries. The access granted to an agent—and the third-party skills it uses—both matter when assessing the safety of autonomous workflows.