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 This may be outdated in 3 to 4 months, or sooner.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.
- Read
- Access
- View only
- Approval
- None needed; check the output
- Move up when
- Output has been right repeatedly
- Draft
- Access
- View plus create drafts
- Approval
- A person sends every draft
- Move up when
- Drafts need little fixing and nothing has surprised you
- 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.
| 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.
- Meta announcement: introducing muse personal ai agent
- xAI documentation: approvals security and privacy
- DataCamp: openai dots
- Cloudflare documentation: features
- AWS announcement: policy amazon bedrock agentcore generally available
- DataCamp: claude duet
- Cloudflare documentation: pricing
- AWS announcement: agentcore identity secrets manager
- Cloudflare documentation: workflows
- 9to5Google: openai dots agent