Skip to main content
All lessons
Lesson 9Social Engineering Defense pathway

AI Agents Need Guardrails

Limit, approve, observe, revoke

Highlighted wordsare glossary terms — tap one for a plain-language definition.

A chatbot mainly gives you information. An can also take actions — reading files, sending messages, changing code, making purchases, or using connected accounts. The more an AI can access and act, the more carefully you must limit its permissions. [1, 2]

Access turns answers into actions

The important difference is authority. A normal chat cannot independently enter your accounts. Risk rises when an agent receives tools, credentials, internet access, or permission to act. In 2026, Anthropic documented four incidents in specially configured cybersecurity evaluations where models reached real external systems after an isolation failure. These were research environments running without ordinary safeguards — not everyday chatbot conversations escaping on their own. [1]

  • Ask what the agent can read, change, send, buy, or delete.
  • Treat every connected account as a real permission, not a convenience toggle.
  • A claim such as 'read-only' does not make broad access necessary or harmless.

Use four guardrails

Use one rule whenever an AI can act: limit, approve, observe, revoke. OWASP identifies excessive autonomy, tool abuse, and high-impact actions without human oversight as agent risks. Its guidance calls for , explicit approval for consequential actions, monitoring, and a way to stop an agent. [2, 3]

  • Limit: give access only to the files, accounts, and permissions needed now.
  • Approve: require confirmation before sending, publishing, purchasing, deleting, deploying, or changing an account.
  • Observe: review logs, sent messages, purchases, and account changes.
  • Revoke: disconnect the agent and remove temporary access when the task ends.

Share the receipt folder — not your digital life

Suppose an agent organizing travel receipts asks for your entire email account, cloud drive, calendar, and payment information. Do not grant broad access. Put copies of the relevant receipts in a separate folder and share only that folder. The safest permission is the smallest one that can complete the task. A 2026 NIST NCCoE draft concept paper identifies least-privilege authorization for software and AI agents as an important open challenge. [4]

  • Do not connect a primary financial account for a receipt-organizing task.
  • Do not let the agent send messages or make purchases when the task needs neither.
  • Use a separate folder or temporary workspace when possible.
  • Remove access after checking the finished work.

Keep high-impact accounts behind stronger boundaries

Some connections can expose money, identity, health information, recovery access, or an entire organization. Do not connect them without narrow permissions, clear approval points, activity records, and a tested way to disconnect access. [2, 3]

  • Banking and payment accounts
  • Password managers and primary email
  • Health portals and cloud-administrator accounts
  • Domain registrars and work systems containing personal or confidential information

Brought to you by NorthBridge

Unfamiliar term? Open the Digital Safety Glossary