Evidence over alarm
← The full AI security framework
AI SECURITY / 02 / AGENTS

Agents.

What can an agent do, and who gave it permission?

The AI that takes action
PICTURE IT THIS WAY

Give a digital assistant a job badge with access to specific rooms. Check what it does inside them, too.

AREA 02 / The AI that takes action

Put boundaries around action.

An agent can mistake a hostile document for an instruction and use its permitted tools in a harmful way. Excessive access can make the damage worse.

01

Name the owner and the job

Record each agent, its purpose, identity, tools, data access, and accountable owner. Include agents supplied by vendors.

02

Enforce limits outside the model

Restrict tools, destinations, credentials, and permissions to the task. Treat instructions found in documents or websites as untrusted input.

03

Check and stop real actions

Require approval for high-impact actions. Record tool calls and outcomes. Test whether revoking access actually stops a running agent.

ASK TO SEE THE EVIDENCE

Test attempts to reach an unapproved system, show the enforced block, and explain the test’s limits. Demonstrate how you revoke the agent’s access.

WHO SHOULD OWN THIS?

AI platform + application owners + security

WHAT TO KEEP IN MIND

A prompt filter or a written rule is not an access boundary. These controls apply to agents you operate or procure; they do not govern an attacker’s agents.

What are these technologies called?

Names you may encounter in a buying conversation. Products often span several areas.

Agent inventory
A record of agents, their owners, purposes, tools, and permissions.
Identity and access controls
Decide who or what may reach a resource, and what actions are allowed.
Tool authorization
Check an agent’s permission before a tool performs an action.
Agent activity monitoring
Record and inspect what agents ask their tools to do and what happens.

Every area depends on clear ownership, policy, and evidence.

See the governance foundation