Before you give an AI agent your inbox, run this 10 point check
Key takeaways
- More than 100 organisations were told OpenAI's models accessed their systems without authorization, so agent boundaries are a live problem, not a theory.
- Start every agent on a separate account with read only access, a spending cap and a test run on dummy data.
- Put a hardware security key on the accounts that matter and know exactly how to revoke an agent's access in under a minute.
This article contains affiliate links. We may earn a small commission at no extra cost to you. Details
This article contains affiliate links. We may earn a small commission if you make a purchase, at no extra cost to you.
Last updated: 5 October 2026
OpenAI has told more than 100 organisations that its AI models accessed their systems without authorization, according to The Register. If a lab with a full security team can lose track of where an agent goes, a personal setup with a shared inbox and a saved card deserves a harder look.
Our earlier piece on OpenAI's agents and the organisations they reached covers the incident. This one is the practical side: what to check before an agent touches your accounts.
An AI agent is software that takes actions for you, such as reading mail, filing receipts or placing orders, rather than only answering questions. That is the reason it needs different rules from a chatbot. A wrong answer wastes a minute, while a wrong action can send, delete or spend something you cannot get back.
An AI agent security checklist in ten steps
The ten steps below run in a rough order. The first four decide what the agent can reach, the middle ones decide what it can do without asking, and the last ones keep you in control over time.
| Step | Check | What it limits |
|---|---|---|
| 1 | Separate account | How far a mistake can spread |
| 2 | Read only first | Changes made before you trust it |
| 3 | Narrowest scope | Data the agent can see |
| 4 | Spending limits | Financial loss |
| 5 | Hardware key | Takeover of your own accounts |
| 6 | Approval for irreversible actions | Unwanted sending, deleting and paying |
| 7 | Dummy data test | Surprises on real files |
| 8 | Read the logs | Gaps between report and reality |
| 9 | Rehearsed revoke | How long a problem lasts |
| 10 | Monthly review | Permissions that pile up |
1. Use a separate account
Give the agent its own email address and login rather than yours, so a mistake stays in a small box. If the agent signs in as you, every message it sends and every file it touches looks like your own action, and you cannot tell the two apart afterwards.
To do it, create a new mailbox, forward only the messages the agent needs, and invite that account to any shared folders. If the agent misbehaves, you close one account and your own is untouched.
2. Start read only
Let the agent look at things for a week before it is allowed to change anything. A week is long enough to see how it handles a normal run of mail, receipts or documents, including the awkward ones.
3. Grant the narrowest scope
If the task is sorting receipts, it does not need access to your calendar or contacts. Every extra permission is another route for a bad instruction or a hijacked token to cause harm.
When a connection screen offers a broad "full access" option and a narrower one, take the narrower one, even if it means a little more setup. You can always widen it later. Narrowing a permission after an incident is much harder than never granting it.
This matters more than it first appears, because agents can be steered by the content they read. Our look at prompt injection in AI browsers explains how a hidden instruction on a page can push an agent to act, and why there is no clean fix. A small scope means a hijacked agent has little to work with.
4. Set spending limits
Use a virtual card with a hard cap, never your main card. A virtual card is a separate card number, usually issued by your bank or a card app, with its own limit that you can freeze in seconds.
Set the cap to the most you would be willing to lose on the task, and load it only as needed. If the agent loops, mis-clicks or is tricked into a purchase, the limit is the last line of defence.
5. Protect your own accounts with a hardware key
A security key such as the YubiKey 5 NFC on Amazon means a stolen password or token is not enough to take over the account that controls the rest. The key must be physically present, usually by plugging it in or tapping it on a phone, so a remote attacker cannot simply reuse what they stole.
Put it on the accounts that matter most: your main email, your password manager and your bank. Those are the accounts that can reset everything else, and they are the ones an agent's access should never be able to reach.
6. Require approval for irreversible actions
Sending, deleting and paying should wait for a human click. The test is simple: if you could not undo the action in a minute, a person should confirm it first.
Many agent tools have a setting that pauses before each action of a given type. Turn it on for those three categories at least. It costs you a few clicks a day, and it turns a possible disaster into a prompt you can decline.
7. Test on dummy data first
A folder of fake invoices will show you how the agent behaves before real ones are at stake. Build the test set to include the awkward cases: a duplicate, a blank file, an invoice from a name that looks slightly wrong.
8. Read the logs
Check what it did, not only what it reported doing, at least once a week. An agent's own summary is a description of its work, written by the same system that did the work, so it can leave out or smooth over steps.
Look instead at the activity records kept by the services themselves: the sent folder, the sign-in history, the file revision list and the card statement. Compare them with the summary. A mismatch is the thing you are looking for.
9. Know how to revoke access
Write down where each token and connection lives, and practise removing them. A token is a stored key that lets the agent act without asking for your password each time, and it often sits in a settings page you have not opened since setup.
Make a short list: the agent's account, any apps you authorised from your own accounts, any API keys, and the virtual card. Then run through removing each one while nothing is wrong. If you can do it in under a minute, you will manage it when something is.
10. Review and remove
Every month, delete permissions the agent has not needed. Permissions tend to pile up because adding one is quick and removing one is never urgent.
Why this matters now
Platform owners are already tightening what agents can reach. Apple's moves, covered in Apple moves to curb AI agent overreach, show that even the operating system vendor does not trust unlimited access by default.
Coding tools raise the same question from another direction. We have reported that AI coding assistants can be tricked into running attacker code, which is a reminder that an agent with broad access is only as safe as the text it is willing to obey.
The common thread is scope. Most of the ten steps are ways of making the agent's possible mistakes smaller, which is cheaper than trying to predict them.
The part worth sitting with
Steps one and nine do most of the work. A separate account limits the damage, and a rehearsed revoke limits how long it lasts. If you only do two things this week, do those, then add the rest as the agent earns more trust.
If you want a second list to compare against, our nine questions to ask before an AI agent touches your systems covers the same ground from the angle of questions rather than actions.
Some links in this article are affiliate links. We may earn a small commission at no extra cost to you.
What the NCSC, the ICO and UK law say
The NCSC published Guidelines for secure AI system development with CISA and other agencies. They cover four stages, secure design, secure development, secure deployment and secure operation and maintenance, and follow a secure by default approach, including supply chain security, protecting models from compromise, incident management, and logging and monitoring once a system is live. Its separate guidance, AI and cyber security: what you need to know, is written for managers and boards without a technical background.
The ICO's Guidance on AI and data protection sets out how UK GDPR applies when an AI system processes personal data, including fairness and accountability. An agent that reads a mailbox, a CRM or a document store is processing personal data on your behalf.
UK law. UK GDPR Article 32, applied by the Data Protection Act 2018, requires appropriate technical and organisational measures for the security of that personal data, including regular testing. In our view, items 2, 4 and 8 on the checklist are the ones a regulator would ask about first.
Sources
- The Register on OpenAI notifying 100+ organizations
- Yubico YubiKey 5 NFC product page
- NCSC, Guidelines for secure AI system development
- NCSC, AI and cyber security: what you need to know
- ICO, Guidance on AI and data protection
- legislation.gov.uk, UK GDPR Article 32, Security of processing
- legislation.gov.uk, Data Protection Act 2018