AI Agent Security
This page describes the connection method and limit rules that apply when AI agents handle protected data.
Limits of existing approaches
Direct opening with user permissions
In a typical connection, an agent opens files directly with the user's permissions.
Delegation with the same scope as the user
When user permissions are delegated unchanged, the agent accesses the same range of data as the user.
Controls dispersed across agents
Building a dedicated integration for each agent splits control points as the number of agents grows.
Existing approaches and AI agent security
Mediated access
AI agents cannot access protected files directly. Their requests go through a mediation step and receive only content within the permitted range.
Even if a file's path is known, the file cannot be opened directly.
Dual limit
The lower of the clearance of the user who delegated the task and the limit the organization set for agents is applied to what an agent can receive.
Even if the user's clearance is higher, documents above the organization's limit are not passed to the agent.
Standard integration protocol
Connections with AI agents use MCP (Model Context Protocol), an open standard.
Agents that support MCP connect under the same controls without separate development.
Agent access records
Agent requests are also recorded in the data access history.
Assurances
Alignment with Zero Trust for non-person entities
The User pillar of Zero Trust covers both human users and non-person entities (NPE).
Access by AI agents is also handled according to this principle.
Try JOLDISK for free
