Quick answer: Before connecting an AI agent to an account, check the developer, permissions, data access, approval settings, and ability to revoke access. Start with low-risk tasks and never give an agent more access than it needs.
AI agents are moving beyond answering questions. They can increasingly interact with email, documents, browsers, databases, cloud services, and business applications. That makes them more useful—but also increases the consequences of a mistake or compromise.
1. Check who made the agent.
Use agents from a known company, established developer, or trusted software provider. Check the official website, privacy policy, support information, app store listing, and update history.
Be cautious if an agent:
- Has no identifiable developer
- Requires unusual permissions
- Uses a suspicious download page
- Promises unlimited access or impossible results
- Has no privacy policy
- Asks you to disable security protections
2. Review every permission.
Read what the agent can access before you approve it. Ask whether it needs access to your entire inbox, cloud drive, contacts, calendar, financial information, or work systems.
A useful rule is:
Give an AI agent the smallest amount of access needed to complete its task.
An agent that summarises one folder does not need access to your entire account.
3. Turn on approval before important actions.
Where possible, require confirmation before the agent can:
- Send messages
- Delete files
- Make purchases
- Change account settings
- Publish content
- Share documents
- Run code
- Transfer money
Automation is convenient, but human approval should remain in the loop for actions that are expensive, public, irreversible, or sensitive.
4. Do not paste sensitive information unnecessarily.
Avoid giving an agent:
- Passwords
- Recovery codes
- Bank details
- Identity documents
- Private medical information
- Confidential business files
- API keys or access tokens
If a tool needs a secret to work, use a secure integration or secret manager rather than pasting the secret into a chat window.
5. Treat web pages and documents as untrusted.
An AI agent may read instructions embedded in a web page, email, document, or attachment. Those instructions may be malicious or designed to manipulate the agent.
Do not assume that an agent can safely distinguish every genuine instruction from an attack. Review important actions yourself, especially when a document asks the agent to reveal information, download software, or change permissions.
6. Test with a low-risk account first.
Before connecting an agent to your main email or work account, test it with:
- A spare account
- Non-sensitive documents
- Read-only permissions
- A limited folder
- A small, reversible task
Observe what it does and what information it stores. Increase access only after you understand its behaviour.
7. Know how to revoke access
Before using an agent, find the account’s connected apps or security settings. Confirm that you can:
- Disconnect the agent.
- Revoke its tokens.
- Change your password if necessary.
- Delete stored data.
- Contact the developer.
Review connected applications regularly and remove tools you no longer use.
Warning signs that an AI agent is unsafe
Stop using the tool if it:
- Acts without asking for approval
- Requests more access than necessary
- Sends data to unknown services
- Cannot explain how data is stored
- Makes unexplained account changes
- Downloads files or software without permission
- Prevents you from revoking access
AI development frameworks and agent components are becoming important security targets because a compromised tool may expose the credentials and services connected to it. 1
The safest way to begin
Start with an agent that can read but not edit. Use it for low-risk tasks such as summarising public documents, organising a small set of notes, or drafting—but not sending—messages.
Only expand its permissions when the benefit is clear and the risk is understood.
Final verdict
AI agents are safest when treated like powerful third-party apps, not like trusted employees. Review permissions, limit access, require approval for important actions, protect sensitive information, and know how to revoke access. Convenience should never come before control.
