Permission ladder
Read-only < draft-only < browser access < send/submit < payment/admin. The higher it climbs, the stronger the approval gate needs to be.
← Home security index
AI
AI agents: if it can act as you, gate it like you
AI can draft, summarise and compare. When it sends, buys, deletes, submits or changes access, it is acting with your identity.
Nigel version
Read-only and draft-only are low drama. Consequential actions need human approval. Secrets do not belong in random tools.
Agent approval gate
Let AI prepare things. Do not let it quietly do things.
A useful home agent can read a bill, compare options or draft a message. The line changes when it sends email, pays money, deletes files or changes access while logged in as you.
Permission posture
The agent can help without becoming the household's unattended hands. Reading and drafting stay separate from doing.
Data boundary
Reading still needs limits: no secrets, client material, medical/legal details or family-sensitive uploads in random tools.
Explain the jargon
Small terms, big consequences
Tap a term for the plain-English version and the practical move. No fake mystique, just the bit that changes what you do at home.
?Prompt injection
Instructions hidden in a page, email or file that try to steer an AI tool away from what you asked it to do.
Do this: Do not let agents act on untrusted content without a review step, especially before sending, buying, deleting or changing access.
?Excessive agency
Giving an AI more ability to act than the task needs: browser sessions, payment access, email sending or admin permissions.
Do this: Default to read-only or draft-only. Add explicit approval gates for consequential actions.
?Sensitive data
Secrets, client material, legal/medical/family information, work files and anything you would not want in a vendor support ticket.
Do this: Use approved tools and keep random AI sites away from sensitive uploads and prompts.
Do this
- Use approved AI tools for work data.
- Keep secrets, client data, legal/medical/family-sensitive material out of random tools.
- Give agents read-only or draft-only access by default.
- Require approval before send/submit/delete/buy/change access.
- Review logs, browser history and connected accounts.
Check
- Can the agent send email?
- Can it make purchases?
- Can it access work data?
- Can it change permissions?
- Is a human approving actions?
Avoid
- Treating an agent like a search box.
- Letting browser agents roam through banking/tax/work accounts.
- Using AI to secretly replace accountable work.
Full guidance
More than a slide title
An approval-gate flow for agent permissions and sensitive actions.
Data discipline
Prompts, uploads, outputs, logs and metadata may exist outside the household. If it would be awkward in a vendor ticket, do not paste it into a random tool.
Human stays accountable
AI can help prepare work. It should not secretly make decisions, submissions or representations on your behalf.
The page is part of the input
A browser agent does not only read the task you typed. It may also read web pages, emails, PDFs, tickets and chat messages. Any of those can contain hostile instructions. That is why untrusted content and action authority are a bad mix: the tool may be looking at a scam page while also holding your logged-in browser.
Good household default
Let AI draft the comparison, write the email, summarise the bill or list the steps. Keep the actual click with a person for sending, buying, deleting, changing account settings, uploading private files or granting access. It is boring, but boring is exactly the point.
Scenario
Helpful bill agent
It wants to log in, compare bills and submit a change.
Better response
- • Let it draft comparison
- • Review manually
- • Submit yourself
Worse habit
Handing it email, payment and submission authority unattended.
AI reads a scam email
A browser agent is asked to organise email and one message says to ignore previous instructions, export data or approve a fake invoice.
Better response
- • Keep the agent draft-only
- • Review messages before any send or payment action
- • Treat instructions inside untrusted content as content, not commands
Worse habit
Letting the agent both read untrusted mail and act on accounts without a human stop.
Random AI site wants a file
A free tool asks for a tax PDF, work document, medical letter or family legal file to summarise.
Better response
- • Use an approved tool for sensitive material
- • Redact what is not needed
- • Check retention and account settings
Worse habit
Uploading the real document because the summary will only take ten seconds.