SecurityApril 27, 2026⏱ 6 min read
The Small Business AI Agent Security Checklist
You don't need to be a security expert to use an AI agent safely. But you do need to check a few boxes. This is the no-jargon checklist for business owners who want to automate responsibly.
Before You Connect Anything
✓
Read the privacy policy
Not the whole thing. Look for three things: (1) Do they sell your data? (2) Do they use your data to train AI models? (3) Can you delete your data when you leave? If the answer to #1 or #2 is yes without your consent, walk away.
✓
Check where your data is stored
Your agent processes business data: emails, calendar entries, customer names. Ask where that data lives. Cloud-hosted in a major provider (AWS, GCP, Azure) is standard. "Our own servers" without further detail is a yellow flag.
✓
Verify encryption in transit and at rest
Your data should be encrypted when it travels between your tools and the agent (in transit) and when it's stored (at rest). This is table stakes in 2026. If the platform can't confirm this, it's behind.
When Connecting Tools (MCP)
✓
Only connect tools your agent actually needs
If your agent handles scheduling and email, it doesn't need access to your banking app. Connect the minimum set of tools for the automations you're running. You can always add more later.
✓
Use read-only access where possible
Many MCP connections let you choose read-only vs. full access. If your agent just needs to check your calendar (not create events), give it read-only. Restrict permissions to what's needed for each specific workflow.
✓
Review connection permissions before approving
When you connect Slack, Gmail, or a CRM via OAuth, you'll see a permissions screen. Actually read it. If a Slack connection asks for admin permissions when it only needs to post messages, that's excessive.
While Your Agent Is Running
✓
Review the activity log weekly
Check what your agent actually did this week. Look for unexpected actions, tools it accessed that you didn't expect, or patterns that don't match your automations. Our
health check guide walks you through this.
✓
Enable "confirm before acting" for sensitive workflows
For workflows that involve money (invoicing, payments) or external communication (sending emails to customers), enable confirmation mode. The agent queues the action and waits for your approval before executing.
✓
Don't share API keys or passwords with your agent through chat
Use the official MCP connection flow or settings panel to provide credentials. Never paste passwords, API keys, or tokens into a conversation. The connection should be handled through secure OAuth or credential management, not copy-paste.
✓
Keep the emergency stop accessible
Know where the kill switch is before you need it. In Agent HQ, it's in Settings. If your agent ever does something unexpected, you should be able to stop it immediately. Test this once so you know it works.
Red Flags to Watch For
⚠ Your agent requests permissions you didn't set up. If your agent tries to access a tool you didn't connect, something is wrong. Disconnect and investigate.
⚠ Activity log shows actions you didn't configure. Every action should trace back to an automation you set up. Unexplained entries need investigation.
⚠ The platform can't explain where your data goes. If support can't give you a clear answer about data storage, processing, and retention, reconsider the platform.
⚠ No option to export or delete your data. You should be able to get your data out and ask for deletion at any time. This isn't just good practice; in many regions it's the law.
OpenClaw's security commitment: All data encrypted in transit (TLS 1.3) and at rest (AES-256). We never sell your data or use it to train models. You can export or delete your data anytime. MCP connections use OAuth with minimal permissions. Full activity logs for every action. Read our
privacy policy for complete details.
Secure by Default
Set up your agent with confidence. Every connection is encrypted, every action is logged.
Get Started Securely