Page 381 - CyberSecurityEssentialsEnglish

FIGURE 8-7: THE GOVERNANCE BOUNDARY AROUND AN INTELLIGENT AGENT

What to notice: An agent sits at the center of a network of tools, data, identity, and
approvals. Every connection needs least privilege, a log, limits, and the ability to stop it.

Source: Prepared by the author, drawing on published guidance on agentic application risks
[6].

TABLE 8-10: PRACTICAL CONTROLS FOR AGENTIC SYSTEMS

Control What it prevents

A separate identity for the
Actions attributed to a human account with wider rights.
agent

A tool allowlist Silent expansion of capability through new integrations.

Least privilege on every data
Reading or writing beyond what the task requires.
source

Approval before irreversible Sending, deleting, purchasing, or paying without a
actions human decision.

Limits on rate, cost, and
Runaway loops and unbounded consumption.
duration

A complete action log An inability to reconstruct what actually happened.

A tested stop control Being unable to halt a session or revoke tokens quickly.

Reviewable memory Persistent state that should not have been retained.

⚠ Warning: capability you cannot explain, stop, or reverse
Do not grant an agent the authority to carry out an action that the organization cannot
explain, stop, or reverse. The less reversible the action, the more the autonomy must be
reduced and the approval and documentation increased.

380