Who Approved This Agent? Rethinking Access, Accountability, and Risk in the Age of AI Agents
Summary: AI agents are being deployed quickly, shared broadly, and granted wide access permissions, and unlike users or service accounts they operate with delegated authority, act autonomously across systems, and often have no clear owner or approver. The article sorts them into personal agents, which inherit a single user's permissions; third-party agents embedded in SaaS platforms, where the vendor owns the agent; and organizational agents shared across teams, which it calls the highest risk and largest blast radius because they carry broad, persistent permissions with no defined lifecycle.
Key Takeaway: Organizational agents can act as authorization bypass paths: a user who cannot directly reach certain data can still trigger an agent that can, and because the agent's credentials are legitimate, traditional access controls raise no alert. Without a map of user to agent to system to action, there is no clear owner to respond or even establish the full blast radius when something goes wrong.
Read the original at The Hacker News ↗