Security for agents: credentials, audits and approval gates

Watch

Chapters

Each chapter starts the class video at that moment.

1. Credentials stay out of the connector layer47:23
2. Shared team bots need their own protocols48:07
3. Four questions before granting access48:28
4. The daily clean sweep48:45
5. Use your subscriptions through the CLI, not API tokens50:01
6. A dedicated CLI bot51:00
7. Don’t manage passwords inside the assistant51:44
8. One segregated password manager52:51
9. Onboard and offboard keys53:17
10. One security agent for every password request53:30
11. A supervisor bot audits every bot nightly54:20
12. A CFO bot gates financial information55:07
13. A Messenger bot gates outbound messages55:37
14. Control a separate laptop from your phone55:58
15. A local open model handles background work57:54
16. Draft first, then ask permission58:22
17. Not built for speed (yet)1:00:19
18. Bots check other bots for context1:01:28
19. You give the final approval1:01:41
20. Q: Does the agent pull passwords itself?1:02:49
21. A: Yes — after a separate code from you1:03:04

From the class

What this module covers

Two member questions prompted this part of class: whether an agent on your own computer can drive another assistant, and how to manage access to a password manager. The questions are trimmed. The video is his answers, including a member-directory app and a chapter forum the agents can work with.

• Credentials don't go through the connector layer. The password manager is reachable only through a security bot, and the agent retrieves a credential only after he enters a separate approval code.
• Shared team bots need their own protocols. Before granting access, ask why it is needed, which bot or person is asking, what it is for, and for how long.
• A daily clean sweep lists files, downloads and documents that may hold confidential information, and asks before anything is deleted.
• Don't manage passwords inside the assistant. Use one segregated password manager for the company or for yourself, so you can onboard and offboard access. He uses 1Password for that.
• One security agent handles every password request. A supervisor bot can check the others against their rules on a schedule.
• A finance bot gates financial information. A messenger bot gates outbound messages: the agent drafts, and you approve. He often edits the draft before it goes.
• From his phone he had an agent take control of a separate laptop that also runs the agent, to send a text and a WhatsApp message. That laptop runs a local open model for background work.
• It is slow, because it works through the laptop's apps one step at a time. If you need a fast reply, this is the wrong tool for that moment.
• The agent checks with his other bots for context before it acts, and he gives the final approval.
• Does the agent pull passwords itself? It can, and before it pulls a password it has to get a separate code from you. That gate is what keeps a credential from being used by accident.
• He showed a member-directory app and a chapter forum in this part of class. A short homework aside on the same subject is not in the recording.

Try it

Put it to work

Secrets stay behind a gate, sends stay behind a gate, and the default is read-only.

Check each provider's terms before running a personal subscription as an always-on background service.

1. Write your security bot's rules: the four access questions, plus a separate approval code before any credential is released.
2. Give each bot one lane, such as research, finance or messages, and write the gate for that lane.
3. Schedule a nightly audit that checks every bot against its rules and reports exceptions.
4. Draft a daily clean-sweep checklist that asks before deleting anything.

You are my security bot. Every credential request from any bot goes through you. Before releasing anything, ask: why is it needed, which bot or person is asking, what is it for, and for how long? Then ask me for my separate approval code. Never show a password in chat.

Every night, check every bot against its security rules. Report any bot that used a credential, shared financial information or sent a message outside its rules. Don't fix anything. Report to me.

Every day at [time], run a clean sweep: list stored files, downloads and documents that may hold confidential information, and propose what to archive or delete. Clear cookies. Flag security tokens due for rotation. Ask me before deleting anything.

Any request to share financial information goes to the finance bot for authorization. Any outbound message goes to the messenger bot as a draft, and nothing is sent without my approval unless I've pre-approved that routine.

Check your work

You're done when

• Every credential your agents use comes through one security bot, with a separate approval step.
• You have written gates for financial information and outbound messages, plus a scheduled audit.

Tools

Mentioned in this module

• 1Password — The password manager he demonstrated: separate vaults, and a credential only after an approval code.
• Composio — Used for other tools. Not used for credentials.

Also on the class guide

Written before class

• Pre-class guide: ground rules
• Pre-class guide: what to tell IT

Between classes

WhatsApp and YouTube

Join the AI Practicum WhatsApp group

Hive Research Institute on YouTube