First BellEric Bakken · AI consulting

← Learn · By Eric Bakken · Updated Nov 2026

Explainer

How to run AI agents securely: a practical checklist

Short answer: Climb a trust ladder: let the agent read, then draft, then act with my okay. Give each person their own account, connect only what the job needs, keep secrets out of chat, decide who may approve, keep a log, and know how to disconnect fast.

Fresh as of 10 Nov 2026. This may be outdated in 3 to 4 months, or sooner.

Start with the trust ladder

I use a trust ladder with three rungs. Read: the agent looks and reports. Draft: it writes, and a person sends. Act with my okay: it does the thing after approval.

Move up one rung per job, not per product. A tool can be on the act rung for one job and the read rung for another. Move a job back down the moment something surprises you.

The ladder is my own principle, not an industry standard. As of 10 Nov 2026, that is how I run it.

The trust ladder
The trust ladderThree-rung ladder: read with view-only access, draft with a person sending, then act only after a named person approves each action. STEP 1Read STEP 2Draft STEP 3Act with my okay
  1. Read
    Access
    View only
    Approval
    None needed; check the output
    Move up when
    Output has been right repeatedly
  2. Draft
    Access
    View plus create drafts
    Approval
    A person sends every draft
    Move up when
    Drafts need little fixing and nothing has surprised you
  3. Act with my okay
    Access
    Can send, book or file
    Approval
    A named person approves each action
    Move up when
    Stay here. Move down if anything surprises you

Accounts and access

Each person uses their own account. Connect only the apps the job needs. You will forget what you connected, so disconnect the leftovers. Meta says you choose which apps connect to Muse and can disconnect any time.

Do not use separate agents as walls. xAI’s docs say Bots on one computer are not a security boundary.

Leave access to your own computer off, or on ask. News outlets report Dots local-computer access is off by default, and xAI says Grok Bot asks every time by default. If you build your own, Cloudflare Access can sit in front of a gateway, and AWS says AgentCore checks policy rules on each tool call.

Security checklist for running an AI agent
Area Check Why
Accounts Each person has their own account Shared logins hide who did what
Accounts Separate agents are not used as walls between clients or secrets xAI says Bots on one computer are not a security boundary
Access Only the apps this job needs are connected Less to reach, less to lose
Access Local computer access is off or set to ask Reported default for Dots; stated default for Grok Bot
Secrets No passwords, cards or IDs typed into chat Use the product’s vault or secure form
Approvals A named person approves money, contracts, deletes A reflex yes is no control
Approvals 'Always allow' rules are few and reviewed They remove the pause
Spend A cap is set, and usage is watched anyway A Grok Bot cap does not stop a run mid-task
Logs You know where the action record is and have read it You need it on a bad day
Exit You know how to disconnect every app Meta says Muse apps can be disconnected any time; find the same switch in your product

Secrets: passwords, cards, IDs

Never paste a password or card number into a chat. Use the product’s own secret handling.

xAI says Grok Bot takes passwords and cards through Vault or Secure Forms, and they are not shown to the model. Meta says Muse checks out through Stripe Link with one-time-use card numbers. News outlets report Duet refuses card numbers and government ID by default.

If you build your own, Cloudflare AI Gateway has DLP scanning for sensitive data patterns, with profiles for financial information and ID numbers. Security researchers reportedly found identity gaps in default AWS AgentCore setups. Set up identity on purpose, not by default.

Approvals: who can say yes

Name who may approve what. Money and contracts go to the owner. Read before you approve. A reflex yes is no control at all.

On Grok Bot, xAI’s docs say that if an Ask first rule and an Allow automatically rule clash, Ask first wins. Background approvals expire after 30 minutes. Be careful with Always allow. xAI says Auto Review should complement, not replace, least privilege.

If you build your own, Cloudflare says Workflows can pause for a person’s approval for hours or days.

Logs: know what it did

Know what the agent did. Find where the product records its actions before you need it, not after. Meta says Muse keeps a complete audit trail. Dots reportedly lets you watch the agent work. If you build your own, Cloudflare AI Gateway logs model calls.

For Duet and Grok Bot I could not confirm from my sources what they record, so check. Read the record on a schedule at first, every morning, say. Then you learn what to watch. No record, no acting.

If something goes wrong

If an agent goes wrong, I stop it first. I disconnect the app, or pause the agent. xAI’s docs say approvals do not undo completed work, so I assume the action happened.

Then I change any password or key it could reach. I read the record and write down what happened while it is fresh. If someone was affected, I tell them plainly.

Then I move that job back down my trust ladder before I try again. This is general practice, not legal advice.

Disclosure: I use Claude and other AI tools, and this site runs on Cloudflare. Where a company’s own page could not be read, I say so and attribute the claim to the news reports.

Quick answers

What is the trust ladder for AI agents?

Three rungs: the agent reads and reports, then drafts for a person to send, then acts with your okay. Move up one rung per job and back down if anything surprises you.

Who should approve an AI agent’s actions?

A named person, and they must read before approving. For money and contracts in a small business, that is the owner.

What should I do if an AI agent makes a mistake?

Disconnect or pause it, assume the action happened, change any passwords it could reach, read the record, tell anyone affected, and move the job back down the ladder.

Keep reading