Free tool

Is your AI agent safe to let loose?

Before an AI assistant or agent touches your email, customers, payments or files, answer sixteen questions. Get a score, the weak areas and what to fix first.

Takes 3 minutes

Sixteen questions in five areas. Nothing you answer leaves your browser.

Choosing a model first? Try the AI model picker.

Access

  • Does the AI use its own account or API key, not a person's login?

  • Can it reach only the systems and folders it needs for this job?

  • Are passwords and API keys kept out of prompts, documents and chat history?

Actions & approvals

  • Does a person approve anything it sends to customers before it goes out?

  • Is it unable to pay, refund, delete or change records without a person approving?

  • Are there limits on how much it can do in one go (messages per hour, records per run)?

  • Is it told to treat emails, web pages and documents as information, never as instructions?

Data & privacy

  • Do you know which personal or customer data it can see, and is that allowed by your privacy terms?

  • Have you checked whether the AI vendor stores your data or uses it for training?

  • Do customers know when they are talking to or receiving messages from an AI?

Monitoring

  • Is every action it takes logged, with what it did and why?

  • Does someone review a sample of its work regularly?

  • Will you be alerted if it fails, loops or does something unusual?

Testing & recovery

  • Was it tested with tricky and wrong inputs before going live?

  • Could you undo its changes (backups or version history of what it can edit)?

  • Can you switch it off in minutes, and does someone know how?

Your score (0 of 16 answered)

0/100

  • Access0%
  • Actions & approvals0%
  • Data & privacy0%
  • Monitoring0%
  • Testing & recovery0%

Want someone to check it properly?

The AI Safety Check maps what your AI can reach, tests the risky paths and gives you the limits to put in place.

Ask about an AI Safety Check →

“Not sure” counts as no: a control you cannot confirm is not protecting you. Any critical gap keeps the result below “Low risk”. This is our working checklist, not a certification.

Why it matters

An AI agent does exactly what it can, not what you meant.

Most problems with AI agents are not clever attacks. They are an agent with too much access, no approval step and no log, doing the wrong thing fast. A few plain controls prevent most of it.

What is an AI agent safety check?
A review of what an AI assistant or agent can reach and do: which accounts it uses, what it can send or change without a person, what data it sees, whether its actions are logged, and whether you could undo them.
Why does Not sure count as No?
Because a control you cannot confirm is not protecting you. If you are not sure the AI is limited to one mailbox, assume it is not until someone checks.
What counts as critical?
Controls that can prevent real damage on their own: separate accounts, least access, approval before sending to customers or moving money, knowing what personal data it sees, logging, and being able to undo its changes. Any critical gap keeps the result below Low risk.
Is this a certification?
No. It is our working checklist from connecting AI to business systems. It is a quick way to see where to look first, not proof that a system is safe.
Want it checked by someone?
The AI Safety Check maps what your AI can reach, tests the risky paths and gives you the limits to put in place.

Connecting AI to your business systems?

Tell us what it will touch. We will check access, approvals and recovery before it goes live, or fix what is already running.

Start a projectFixed-price checks